xref: /freebsd/sys/kern/vfs_aio.c (revision d06955f9bdb1416d9196043ed781f9b36dae9adc)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. John S. Dyson's name may not be used to endorse or promote products
12  *    derived from this software without specific prior written permission.
13  *
14  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
15  * bad that happens because of using this software isn't the responsibility
16  * of the author.  This software is distributed AS-IS.
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
25 
26 #include "opt_compat.h"
27 
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/malloc.h>
31 #include <sys/bio.h>
32 #include <sys/buf.h>
33 #include <sys/capsicum.h>
34 #include <sys/eventhandler.h>
35 #include <sys/sysproto.h>
36 #include <sys/filedesc.h>
37 #include <sys/kernel.h>
38 #include <sys/module.h>
39 #include <sys/kthread.h>
40 #include <sys/fcntl.h>
41 #include <sys/file.h>
42 #include <sys/limits.h>
43 #include <sys/lock.h>
44 #include <sys/mutex.h>
45 #include <sys/unistd.h>
46 #include <sys/posix4.h>
47 #include <sys/proc.h>
48 #include <sys/resourcevar.h>
49 #include <sys/signalvar.h>
50 #include <sys/syscallsubr.h>
51 #include <sys/protosw.h>
52 #include <sys/rwlock.h>
53 #include <sys/sema.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/syscall.h>
57 #include <sys/sysent.h>
58 #include <sys/sysctl.h>
59 #include <sys/syslog.h>
60 #include <sys/sx.h>
61 #include <sys/taskqueue.h>
62 #include <sys/vnode.h>
63 #include <sys/conf.h>
64 #include <sys/event.h>
65 #include <sys/mount.h>
66 #include <geom/geom.h>
67 
68 #include <machine/atomic.h>
69 
70 #include <vm/vm.h>
71 #include <vm/vm_page.h>
72 #include <vm/vm_extern.h>
73 #include <vm/pmap.h>
74 #include <vm/vm_map.h>
75 #include <vm/vm_object.h>
76 #include <vm/uma.h>
77 #include <sys/aio.h>
78 
79 /*
80  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
81  * overflow. (XXX will be removed soon.)
82  */
83 static u_long jobrefid;
84 
85 /*
86  * Counter for aio_fsync.
87  */
88 static uint64_t jobseqno;
89 
90 #ifndef MAX_AIO_PER_PROC
91 #define MAX_AIO_PER_PROC	32
92 #endif
93 
94 #ifndef MAX_AIO_QUEUE_PER_PROC
95 #define MAX_AIO_QUEUE_PER_PROC	256
96 #endif
97 
98 #ifndef MAX_AIO_QUEUE
99 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
100 #endif
101 
102 #ifndef MAX_BUF_AIO
103 #define MAX_BUF_AIO		16
104 #endif
105 
106 FEATURE(aio, "Asynchronous I/O");
107 SYSCTL_DECL(_p1003_1b);
108 
109 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
110 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
111 
112 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
113     "Async IO management");
114 
115 static int enable_aio_unsafe = 0;
116 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
117     "Permit asynchronous IO on all file types, not just known-safe types");
118 
119 static unsigned int unsafe_warningcnt = 1;
120 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
121     &unsafe_warningcnt, 0,
122     "Warnings that will be triggered upon failed IO requests on unsafe files");
123 
124 static int max_aio_procs = MAX_AIO_PROCS;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
126     "Maximum number of kernel processes to use for handling async IO ");
127 
128 static int num_aio_procs = 0;
129 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
130     "Number of presently active kernel processes for async IO");
131 
132 /*
133  * The code will adjust the actual number of AIO processes towards this
134  * number when it gets a chance.
135  */
136 static int target_aio_procs = TARGET_AIO_PROCS;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
138     0,
139     "Preferred number of ready kernel processes for async IO");
140 
141 static int max_queue_count = MAX_AIO_QUEUE;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
143     "Maximum number of aio requests to queue, globally");
144 
145 static int num_queue_count = 0;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
147     "Number of queued aio requests");
148 
149 static int num_buf_aio = 0;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
151     "Number of aio requests presently handled by the buf subsystem");
152 
153 /* Number of async I/O processes in the process of being started */
154 /* XXX This should be local to aio_aqueue() */
155 static int num_aio_resv_start = 0;
156 
157 static int aiod_lifetime;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
159     "Maximum lifetime for idle aiod");
160 
161 static int max_aio_per_proc = MAX_AIO_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
163     0,
164     "Maximum active aio requests per process (stored in the process)");
165 
166 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
168     &max_aio_queue_per_proc, 0,
169     "Maximum queued aio requests per process (stored in the process)");
170 
171 static int max_buf_aio = MAX_BUF_AIO;
172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
173     "Maximum buf aio requests per process (stored in the process)");
174 
175 /*
176  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
177  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
178  * vfs.aio.aio_listio_max.
179  */
180 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
181     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
182     0, "Maximum aio requests for a single lio_listio call");
183 
184 #ifdef COMPAT_FREEBSD6
185 typedef struct oaiocb {
186 	int	aio_fildes;		/* File descriptor */
187 	off_t	aio_offset;		/* File offset for I/O */
188 	volatile void *aio_buf;         /* I/O buffer in process space */
189 	size_t	aio_nbytes;		/* Number of bytes for I/O */
190 	struct	osigevent aio_sigevent;	/* Signal to deliver */
191 	int	aio_lio_opcode;		/* LIO opcode */
192 	int	aio_reqprio;		/* Request priority -- ignored */
193 	struct	__aiocb_private	_aiocb_private;
194 } oaiocb_t;
195 #endif
196 
197 /*
198  * Below is a key of locks used to protect each member of struct kaiocb
199  * aioliojob and kaioinfo and any backends.
200  *
201  * * - need not protected
202  * a - locked by kaioinfo lock
203  * b - locked by backend lock, the backend lock can be null in some cases,
204  *     for example, BIO belongs to this type, in this case, proc lock is
205  *     reused.
206  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
207  */
208 
209 /*
210  * If the routine that services an AIO request blocks while running in an
211  * AIO kernel process it can starve other I/O requests.  BIO requests
212  * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
213  * processes at all.  Socket I/O requests use a separate pool of
214  * kprocs and also force non-blocking I/O.  Other file I/O requests
215  * use the generic fo_read/fo_write operations which can block.  The
216  * fsync and mlock operations can also block while executing.  Ideally
217  * none of these requests would block while executing.
218  *
219  * Note that the service routines cannot toggle O_NONBLOCK in the file
220  * structure directly while handling a request due to races with
221  * userland threads.
222  */
223 
224 /* jobflags */
225 #define	KAIOCB_QUEUEING		0x01
226 #define	KAIOCB_CANCELLED	0x02
227 #define	KAIOCB_CANCELLING	0x04
228 #define	KAIOCB_CHECKSYNC	0x08
229 #define	KAIOCB_CLEARED		0x10
230 #define	KAIOCB_FINISHED		0x20
231 
232 /*
233  * AIO process info
234  */
235 #define AIOP_FREE	0x1			/* proc on free queue */
236 
237 struct aioproc {
238 	int	aioprocflags;			/* (c) AIO proc flags */
239 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
240 	struct	proc *aioproc;			/* (*) the AIO proc */
241 };
242 
243 /*
244  * data-structure for lio signal management
245  */
246 struct aioliojob {
247 	int	lioj_flags;			/* (a) listio flags */
248 	int	lioj_count;			/* (a) listio flags */
249 	int	lioj_finished_count;		/* (a) listio flags */
250 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
251 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
252 	struct	knlist klist;			/* (a) list of knotes */
253 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
254 };
255 
256 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
257 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
258 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
259 
260 /*
261  * per process aio data structure
262  */
263 struct kaioinfo {
264 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
265 	int	kaio_flags;		/* (a) per process kaio flags */
266 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
267 	int	kaio_active_count;	/* (c) number of currently used AIOs */
268 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
269 	int	kaio_count;		/* (a) size of AIO queue */
270 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
271 	int	kaio_buffer_count;	/* (a) number of physio buffers */
272 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
273 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
274 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
276 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
277 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
278 	struct	task kaio_task;		/* (*) task to kick aio processes */
279 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
280 };
281 
282 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
286 
287 #define KAIO_RUNDOWN	0x1	/* process is being run down */
288 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
289 
290 /*
291  * Operations used to interact with userland aio control blocks.
292  * Different ABIs provide their own operations.
293  */
294 struct aiocb_ops {
295 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
296 	long	(*fetch_status)(struct aiocb *ujob);
297 	long	(*fetch_error)(struct aiocb *ujob);
298 	int	(*store_status)(struct aiocb *ujob, long status);
299 	int	(*store_error)(struct aiocb *ujob, long error);
300 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303 
304 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309 
310 void		aio_init_aioinfo(struct proc *p);
311 static int	aio_onceonly(void);
312 static int	aio_free_entry(struct kaiocb *job);
313 static void	aio_process_rw(struct kaiocb *job);
314 static void	aio_process_sync(struct kaiocb *job);
315 static void	aio_process_mlock(struct kaiocb *job);
316 static void	aio_schedule_fsync(void *context, int pending);
317 static int	aio_newproc(int *);
318 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
319 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void	aio_physwakeup(struct bio *bp);
322 static void	aio_proc_rundown(void *arg, struct proc *p);
323 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
324 		    struct image_params *imgp);
325 static int	aio_qphysio(struct proc *p, struct kaiocb *job);
326 static void	aio_daemon(void *param);
327 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int	aio_kick(struct proc *userp);
330 static void	aio_kick_nowait(struct proc *userp);
331 static void	aio_kick_helper(void *context, int pending);
332 static int	filt_aioattach(struct knote *kn);
333 static void	filt_aiodetach(struct knote *kn);
334 static int	filt_aio(struct knote *kn, long hint);
335 static int	filt_lioattach(struct knote *kn);
336 static void	filt_liodetach(struct knote *kn);
337 static int	filt_lio(struct knote *kn, long hint);
338 
339 /*
340  * Zones for:
341  * 	kaio	Per process async io info
342  *	aiop	async io process data
343  *	aiocb	async io jobs
344  *	aiolio	list io jobs
345  */
346 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
347 
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
350 	.f_isfd = 0,
351 	.f_attach = filt_aioattach,
352 	.f_detach = filt_aiodetach,
353 	.f_event = filt_aio,
354 };
355 static struct filterops lio_filtops = {
356 	.f_isfd = 0,
357 	.f_attach = filt_lioattach,
358 	.f_detach = filt_liodetach,
359 	.f_event = filt_lio
360 };
361 
362 static eventhandler_tag exit_tag, exec_tag;
363 
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
365 
366 /*
367  * Main operations function for use as a kernel module.
368  */
369 static int
370 aio_modload(struct module *module, int cmd, void *arg)
371 {
372 	int error = 0;
373 
374 	switch (cmd) {
375 	case MOD_LOAD:
376 		aio_onceonly();
377 		break;
378 	case MOD_SHUTDOWN:
379 		break;
380 	default:
381 		error = EOPNOTSUPP;
382 		break;
383 	}
384 	return (error);
385 }
386 
387 static moduledata_t aio_mod = {
388 	"aio",
389 	&aio_modload,
390 	NULL
391 };
392 
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
395 
396 /*
397  * Startup initialization
398  */
399 static int
400 aio_onceonly(void)
401 {
402 
403 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 	    EVENTHANDLER_PRI_ANY);
405 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 	    NULL, EVENTHANDLER_PRI_ANY);
407 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 	TAILQ_INIT(&aio_freeproc);
410 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 	TAILQ_INIT(&aio_jobs);
413 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
416 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
417 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
418 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
419 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
420 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
421 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
422 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
423 	jobrefid = 1;
424 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
425 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
426 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
427 
428 	return (0);
429 }
430 
431 /*
432  * Init the per-process aioinfo structure.  The aioinfo limits are set
433  * per-process for user limit (resource) management.
434  */
435 void
436 aio_init_aioinfo(struct proc *p)
437 {
438 	struct kaioinfo *ki;
439 
440 	ki = uma_zalloc(kaio_zone, M_WAITOK);
441 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
442 	ki->kaio_flags = 0;
443 	ki->kaio_maxactive_count = max_aio_per_proc;
444 	ki->kaio_active_count = 0;
445 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
446 	ki->kaio_count = 0;
447 	ki->kaio_ballowed_count = max_buf_aio;
448 	ki->kaio_buffer_count = 0;
449 	TAILQ_INIT(&ki->kaio_all);
450 	TAILQ_INIT(&ki->kaio_done);
451 	TAILQ_INIT(&ki->kaio_jobqueue);
452 	TAILQ_INIT(&ki->kaio_liojoblist);
453 	TAILQ_INIT(&ki->kaio_syncqueue);
454 	TAILQ_INIT(&ki->kaio_syncready);
455 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
456 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
457 	PROC_LOCK(p);
458 	if (p->p_aioinfo == NULL) {
459 		p->p_aioinfo = ki;
460 		PROC_UNLOCK(p);
461 	} else {
462 		PROC_UNLOCK(p);
463 		mtx_destroy(&ki->kaio_mtx);
464 		uma_zfree(kaio_zone, ki);
465 	}
466 
467 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
468 		aio_newproc(NULL);
469 }
470 
471 static int
472 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
473 {
474 	struct thread *td;
475 	int error;
476 
477 	error = sigev_findtd(p, sigev, &td);
478 	if (error)
479 		return (error);
480 	if (!KSI_ONQ(ksi)) {
481 		ksiginfo_set_sigev(ksi, sigev);
482 		ksi->ksi_code = SI_ASYNCIO;
483 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
484 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
485 	}
486 	PROC_UNLOCK(p);
487 	return (error);
488 }
489 
490 /*
491  * Free a job entry.  Wait for completion if it is currently active, but don't
492  * delay forever.  If we delay, we return a flag that says that we have to
493  * restart the queue scan.
494  */
495 static int
496 aio_free_entry(struct kaiocb *job)
497 {
498 	struct kaioinfo *ki;
499 	struct aioliojob *lj;
500 	struct proc *p;
501 
502 	p = job->userproc;
503 	MPASS(curproc == p);
504 	ki = p->p_aioinfo;
505 	MPASS(ki != NULL);
506 
507 	AIO_LOCK_ASSERT(ki, MA_OWNED);
508 	MPASS(job->jobflags & KAIOCB_FINISHED);
509 
510 	atomic_subtract_int(&num_queue_count, 1);
511 
512 	ki->kaio_count--;
513 	MPASS(ki->kaio_count >= 0);
514 
515 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
516 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
517 
518 	lj = job->lio;
519 	if (lj) {
520 		lj->lioj_count--;
521 		lj->lioj_finished_count--;
522 
523 		if (lj->lioj_count == 0) {
524 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
525 			/* lio is going away, we need to destroy any knotes */
526 			knlist_delete(&lj->klist, curthread, 1);
527 			PROC_LOCK(p);
528 			sigqueue_take(&lj->lioj_ksi);
529 			PROC_UNLOCK(p);
530 			uma_zfree(aiolio_zone, lj);
531 		}
532 	}
533 
534 	/* job is going away, we need to destroy any knotes */
535 	knlist_delete(&job->klist, curthread, 1);
536 	PROC_LOCK(p);
537 	sigqueue_take(&job->ksi);
538 	PROC_UNLOCK(p);
539 
540 	AIO_UNLOCK(ki);
541 
542 	/*
543 	 * The thread argument here is used to find the owning process
544 	 * and is also passed to fo_close() which may pass it to various
545 	 * places such as devsw close() routines.  Because of that, we
546 	 * need a thread pointer from the process owning the job that is
547 	 * persistent and won't disappear out from under us or move to
548 	 * another process.
549 	 *
550 	 * Currently, all the callers of this function call it to remove
551 	 * a kaiocb from the current process' job list either via a
552 	 * syscall or due to the current process calling exit() or
553 	 * execve().  Thus, we know that p == curproc.  We also know that
554 	 * curthread can't exit since we are curthread.
555 	 *
556 	 * Therefore, we use curthread as the thread to pass to
557 	 * knlist_delete().  This does mean that it is possible for the
558 	 * thread pointer at close time to differ from the thread pointer
559 	 * at open time, but this is already true of file descriptors in
560 	 * a multithreaded process.
561 	 */
562 	if (job->fd_file)
563 		fdrop(job->fd_file, curthread);
564 	crfree(job->cred);
565 	uma_zfree(aiocb_zone, job);
566 	AIO_LOCK(ki);
567 
568 	return (0);
569 }
570 
571 static void
572 aio_proc_rundown_exec(void *arg, struct proc *p,
573     struct image_params *imgp __unused)
574 {
575    	aio_proc_rundown(arg, p);
576 }
577 
578 static int
579 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
580 {
581 	aio_cancel_fn_t *func;
582 	int cancelled;
583 
584 	AIO_LOCK_ASSERT(ki, MA_OWNED);
585 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
586 		return (0);
587 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
588 	job->jobflags |= KAIOCB_CANCELLED;
589 
590 	func = job->cancel_fn;
591 
592 	/*
593 	 * If there is no cancel routine, just leave the job marked as
594 	 * cancelled.  The job should be in active use by a caller who
595 	 * should complete it normally or when it fails to install a
596 	 * cancel routine.
597 	 */
598 	if (func == NULL)
599 		return (0);
600 
601 	/*
602 	 * Set the CANCELLING flag so that aio_complete() will defer
603 	 * completions of this job.  This prevents the job from being
604 	 * freed out from under the cancel callback.  After the
605 	 * callback any deferred completion (whether from the callback
606 	 * or any other source) will be completed.
607 	 */
608 	job->jobflags |= KAIOCB_CANCELLING;
609 	AIO_UNLOCK(ki);
610 	func(job);
611 	AIO_LOCK(ki);
612 	job->jobflags &= ~KAIOCB_CANCELLING;
613 	if (job->jobflags & KAIOCB_FINISHED) {
614 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
615 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
616 		aio_bio_done_notify(p, job);
617 	} else {
618 		/*
619 		 * The cancel callback might have scheduled an
620 		 * operation to cancel this request, but it is
621 		 * only counted as cancelled if the request is
622 		 * cancelled when the callback returns.
623 		 */
624 		cancelled = 0;
625 	}
626 	return (cancelled);
627 }
628 
629 /*
630  * Rundown the jobs for a given process.
631  */
632 static void
633 aio_proc_rundown(void *arg, struct proc *p)
634 {
635 	struct kaioinfo *ki;
636 	struct aioliojob *lj;
637 	struct kaiocb *job, *jobn;
638 
639 	KASSERT(curthread->td_proc == p,
640 	    ("%s: called on non-curproc", __func__));
641 	ki = p->p_aioinfo;
642 	if (ki == NULL)
643 		return;
644 
645 	AIO_LOCK(ki);
646 	ki->kaio_flags |= KAIO_RUNDOWN;
647 
648 restart:
649 
650 	/*
651 	 * Try to cancel all pending requests. This code simulates
652 	 * aio_cancel on all pending I/O requests.
653 	 */
654 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
655 		aio_cancel_job(p, ki, job);
656 	}
657 
658 	/* Wait for all running I/O to be finished */
659 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
660 		ki->kaio_flags |= KAIO_WAKEUP;
661 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
662 		goto restart;
663 	}
664 
665 	/* Free all completed I/O requests. */
666 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
667 		aio_free_entry(job);
668 
669 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
670 		if (lj->lioj_count == 0) {
671 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
672 			knlist_delete(&lj->klist, curthread, 1);
673 			PROC_LOCK(p);
674 			sigqueue_take(&lj->lioj_ksi);
675 			PROC_UNLOCK(p);
676 			uma_zfree(aiolio_zone, lj);
677 		} else {
678 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
679 			    lj->lioj_count, lj->lioj_finished_count);
680 		}
681 	}
682 	AIO_UNLOCK(ki);
683 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
684 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
685 	mtx_destroy(&ki->kaio_mtx);
686 	uma_zfree(kaio_zone, ki);
687 	p->p_aioinfo = NULL;
688 }
689 
690 /*
691  * Select a job to run (called by an AIO daemon).
692  */
693 static struct kaiocb *
694 aio_selectjob(struct aioproc *aiop)
695 {
696 	struct kaiocb *job;
697 	struct kaioinfo *ki;
698 	struct proc *userp;
699 
700 	mtx_assert(&aio_job_mtx, MA_OWNED);
701 restart:
702 	TAILQ_FOREACH(job, &aio_jobs, list) {
703 		userp = job->userproc;
704 		ki = userp->p_aioinfo;
705 
706 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
707 			TAILQ_REMOVE(&aio_jobs, job, list);
708 			if (!aio_clear_cancel_function(job))
709 				goto restart;
710 
711 			/* Account for currently active jobs. */
712 			ki->kaio_active_count++;
713 			break;
714 		}
715 	}
716 	return (job);
717 }
718 
719 /*
720  * Move all data to a permanent storage device.  This code
721  * simulates the fsync syscall.
722  */
723 static int
724 aio_fsync_vnode(struct thread *td, struct vnode *vp)
725 {
726 	struct mount *mp;
727 	int error;
728 
729 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
730 		goto drop;
731 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
732 	if (vp->v_object != NULL) {
733 		VM_OBJECT_WLOCK(vp->v_object);
734 		vm_object_page_clean(vp->v_object, 0, 0, 0);
735 		VM_OBJECT_WUNLOCK(vp->v_object);
736 	}
737 	error = VOP_FSYNC(vp, MNT_WAIT, td);
738 
739 	VOP_UNLOCK(vp, 0);
740 	vn_finished_write(mp);
741 drop:
742 	return (error);
743 }
744 
745 /*
746  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
747  * does the I/O request for the non-physio version of the operations.  The
748  * normal vn operations are used, and this code should work in all instances
749  * for every type of file, including pipes, sockets, fifos, and regular files.
750  *
751  * XXX I don't think it works well for socket, pipe, and fifo.
752  */
753 static void
754 aio_process_rw(struct kaiocb *job)
755 {
756 	struct ucred *td_savedcred;
757 	struct thread *td;
758 	struct aiocb *cb;
759 	struct file *fp;
760 	struct uio auio;
761 	struct iovec aiov;
762 	ssize_t cnt;
763 	long msgsnd_st, msgsnd_end;
764 	long msgrcv_st, msgrcv_end;
765 	long oublock_st, oublock_end;
766 	long inblock_st, inblock_end;
767 	int error;
768 
769 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
770 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
771 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
772 
773 	aio_switch_vmspace(job);
774 	td = curthread;
775 	td_savedcred = td->td_ucred;
776 	td->td_ucred = job->cred;
777 	cb = &job->uaiocb;
778 	fp = job->fd_file;
779 
780 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
781 	aiov.iov_len = cb->aio_nbytes;
782 
783 	auio.uio_iov = &aiov;
784 	auio.uio_iovcnt = 1;
785 	auio.uio_offset = cb->aio_offset;
786 	auio.uio_resid = cb->aio_nbytes;
787 	cnt = cb->aio_nbytes;
788 	auio.uio_segflg = UIO_USERSPACE;
789 	auio.uio_td = td;
790 
791 	msgrcv_st = td->td_ru.ru_msgrcv;
792 	msgsnd_st = td->td_ru.ru_msgsnd;
793 	inblock_st = td->td_ru.ru_inblock;
794 	oublock_st = td->td_ru.ru_oublock;
795 
796 	/*
797 	 * aio_aqueue() acquires a reference to the file that is
798 	 * released in aio_free_entry().
799 	 */
800 	if (cb->aio_lio_opcode == LIO_READ) {
801 		auio.uio_rw = UIO_READ;
802 		if (auio.uio_resid == 0)
803 			error = 0;
804 		else
805 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
806 	} else {
807 		if (fp->f_type == DTYPE_VNODE)
808 			bwillwrite();
809 		auio.uio_rw = UIO_WRITE;
810 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
811 	}
812 	msgrcv_end = td->td_ru.ru_msgrcv;
813 	msgsnd_end = td->td_ru.ru_msgsnd;
814 	inblock_end = td->td_ru.ru_inblock;
815 	oublock_end = td->td_ru.ru_oublock;
816 
817 	job->msgrcv = msgrcv_end - msgrcv_st;
818 	job->msgsnd = msgsnd_end - msgsnd_st;
819 	job->inblock = inblock_end - inblock_st;
820 	job->outblock = oublock_end - oublock_st;
821 
822 	if ((error) && (auio.uio_resid != cnt)) {
823 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
824 			error = 0;
825 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
826 			PROC_LOCK(job->userproc);
827 			kern_psignal(job->userproc, SIGPIPE);
828 			PROC_UNLOCK(job->userproc);
829 		}
830 	}
831 
832 	cnt -= auio.uio_resid;
833 	td->td_ucred = td_savedcred;
834 	if (error)
835 		aio_complete(job, -1, error);
836 	else
837 		aio_complete(job, cnt, 0);
838 }
839 
840 static void
841 aio_process_sync(struct kaiocb *job)
842 {
843 	struct thread *td = curthread;
844 	struct ucred *td_savedcred = td->td_ucred;
845 	struct file *fp = job->fd_file;
846 	int error = 0;
847 
848 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
849 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
850 
851 	td->td_ucred = job->cred;
852 	if (fp->f_vnode != NULL)
853 		error = aio_fsync_vnode(td, fp->f_vnode);
854 	td->td_ucred = td_savedcred;
855 	if (error)
856 		aio_complete(job, -1, error);
857 	else
858 		aio_complete(job, 0, 0);
859 }
860 
861 static void
862 aio_process_mlock(struct kaiocb *job)
863 {
864 	struct aiocb *cb = &job->uaiocb;
865 	int error;
866 
867 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
868 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
869 
870 	aio_switch_vmspace(job);
871 	error = kern_mlock(job->userproc, job->cred,
872 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
873 	aio_complete(job, error != 0 ? -1 : 0, error);
874 }
875 
876 static void
877 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
878 {
879 	struct aioliojob *lj;
880 	struct kaioinfo *ki;
881 	struct kaiocb *sjob, *sjobn;
882 	int lj_done;
883 	bool schedule_fsync;
884 
885 	ki = userp->p_aioinfo;
886 	AIO_LOCK_ASSERT(ki, MA_OWNED);
887 	lj = job->lio;
888 	lj_done = 0;
889 	if (lj) {
890 		lj->lioj_finished_count++;
891 		if (lj->lioj_count == lj->lioj_finished_count)
892 			lj_done = 1;
893 	}
894 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
895 	MPASS(job->jobflags & KAIOCB_FINISHED);
896 
897 	if (ki->kaio_flags & KAIO_RUNDOWN)
898 		goto notification_done;
899 
900 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
901 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
902 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
903 
904 	KNOTE_LOCKED(&job->klist, 1);
905 
906 	if (lj_done) {
907 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
908 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
909 			KNOTE_LOCKED(&lj->klist, 1);
910 		}
911 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
912 		    == LIOJ_SIGNAL
913 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
914 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
915 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
916 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
917 		}
918 	}
919 
920 notification_done:
921 	if (job->jobflags & KAIOCB_CHECKSYNC) {
922 		schedule_fsync = false;
923 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
924 			if (job->fd_file != sjob->fd_file ||
925 			    job->seqno >= sjob->seqno)
926 				continue;
927 			if (--sjob->pending > 0)
928 				continue;
929 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
930 			if (!aio_clear_cancel_function_locked(sjob))
931 				continue;
932 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
933 			schedule_fsync = true;
934 		}
935 		if (schedule_fsync)
936 			taskqueue_enqueue(taskqueue_aiod_kick,
937 			    &ki->kaio_sync_task);
938 	}
939 	if (ki->kaio_flags & KAIO_WAKEUP) {
940 		ki->kaio_flags &= ~KAIO_WAKEUP;
941 		wakeup(&userp->p_aioinfo);
942 	}
943 }
944 
945 static void
946 aio_schedule_fsync(void *context, int pending)
947 {
948 	struct kaioinfo *ki;
949 	struct kaiocb *job;
950 
951 	ki = context;
952 	AIO_LOCK(ki);
953 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
954 		job = TAILQ_FIRST(&ki->kaio_syncready);
955 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
956 		AIO_UNLOCK(ki);
957 		aio_schedule(job, aio_process_sync);
958 		AIO_LOCK(ki);
959 	}
960 	AIO_UNLOCK(ki);
961 }
962 
963 bool
964 aio_cancel_cleared(struct kaiocb *job)
965 {
966 
967 	/*
968 	 * The caller should hold the same queue lock held when
969 	 * aio_clear_cancel_function() was called and set this flag
970 	 * ensuring this check sees an up-to-date value.  However,
971 	 * there is no way to assert that.
972 	 */
973 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
974 }
975 
976 static bool
977 aio_clear_cancel_function_locked(struct kaiocb *job)
978 {
979 
980 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
981 	MPASS(job->cancel_fn != NULL);
982 	if (job->jobflags & KAIOCB_CANCELLING) {
983 		job->jobflags |= KAIOCB_CLEARED;
984 		return (false);
985 	}
986 	job->cancel_fn = NULL;
987 	return (true);
988 }
989 
990 bool
991 aio_clear_cancel_function(struct kaiocb *job)
992 {
993 	struct kaioinfo *ki;
994 	bool ret;
995 
996 	ki = job->userproc->p_aioinfo;
997 	AIO_LOCK(ki);
998 	ret = aio_clear_cancel_function_locked(job);
999 	AIO_UNLOCK(ki);
1000 	return (ret);
1001 }
1002 
1003 static bool
1004 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1005 {
1006 
1007 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1008 	if (job->jobflags & KAIOCB_CANCELLED)
1009 		return (false);
1010 	job->cancel_fn = func;
1011 	return (true);
1012 }
1013 
1014 bool
1015 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1016 {
1017 	struct kaioinfo *ki;
1018 	bool ret;
1019 
1020 	ki = job->userproc->p_aioinfo;
1021 	AIO_LOCK(ki);
1022 	ret = aio_set_cancel_function_locked(job, func);
1023 	AIO_UNLOCK(ki);
1024 	return (ret);
1025 }
1026 
1027 void
1028 aio_complete(struct kaiocb *job, long status, int error)
1029 {
1030 	struct kaioinfo *ki;
1031 	struct proc *userp;
1032 
1033 	job->uaiocb._aiocb_private.error = error;
1034 	job->uaiocb._aiocb_private.status = status;
1035 
1036 	userp = job->userproc;
1037 	ki = userp->p_aioinfo;
1038 
1039 	AIO_LOCK(ki);
1040 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1041 	    ("duplicate aio_complete"));
1042 	job->jobflags |= KAIOCB_FINISHED;
1043 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1044 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1045 		aio_bio_done_notify(userp, job);
1046 	}
1047 	AIO_UNLOCK(ki);
1048 }
1049 
1050 void
1051 aio_cancel(struct kaiocb *job)
1052 {
1053 
1054 	aio_complete(job, -1, ECANCELED);
1055 }
1056 
1057 void
1058 aio_switch_vmspace(struct kaiocb *job)
1059 {
1060 
1061 	vmspace_switch_aio(job->userproc->p_vmspace);
1062 }
1063 
1064 /*
1065  * The AIO daemon, most of the actual work is done in aio_process_*,
1066  * but the setup (and address space mgmt) is done in this routine.
1067  */
1068 static void
1069 aio_daemon(void *_id)
1070 {
1071 	struct kaiocb *job;
1072 	struct aioproc *aiop;
1073 	struct kaioinfo *ki;
1074 	struct proc *p;
1075 	struct vmspace *myvm;
1076 	struct thread *td = curthread;
1077 	int id = (intptr_t)_id;
1078 
1079 	/*
1080 	 * Grab an extra reference on the daemon's vmspace so that it
1081 	 * doesn't get freed by jobs that switch to a different
1082 	 * vmspace.
1083 	 */
1084 	p = td->td_proc;
1085 	myvm = vmspace_acquire_ref(p);
1086 
1087 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1088 
1089 	/*
1090 	 * Allocate and ready the aio control info.  There is one aiop structure
1091 	 * per daemon.
1092 	 */
1093 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1094 	aiop->aioproc = p;
1095 	aiop->aioprocflags = 0;
1096 
1097 	/*
1098 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1099 	 * and creating too many daemons.)
1100 	 */
1101 	sema_post(&aio_newproc_sem);
1102 
1103 	mtx_lock(&aio_job_mtx);
1104 	for (;;) {
1105 		/*
1106 		 * Take daemon off of free queue
1107 		 */
1108 		if (aiop->aioprocflags & AIOP_FREE) {
1109 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1110 			aiop->aioprocflags &= ~AIOP_FREE;
1111 		}
1112 
1113 		/*
1114 		 * Check for jobs.
1115 		 */
1116 		while ((job = aio_selectjob(aiop)) != NULL) {
1117 			mtx_unlock(&aio_job_mtx);
1118 
1119 			ki = job->userproc->p_aioinfo;
1120 			job->handle_fn(job);
1121 
1122 			mtx_lock(&aio_job_mtx);
1123 			/* Decrement the active job count. */
1124 			ki->kaio_active_count--;
1125 		}
1126 
1127 		/*
1128 		 * Disconnect from user address space.
1129 		 */
1130 		if (p->p_vmspace != myvm) {
1131 			mtx_unlock(&aio_job_mtx);
1132 			vmspace_switch_aio(myvm);
1133 			mtx_lock(&aio_job_mtx);
1134 			/*
1135 			 * We have to restart to avoid race, we only sleep if
1136 			 * no job can be selected.
1137 			 */
1138 			continue;
1139 		}
1140 
1141 		mtx_assert(&aio_job_mtx, MA_OWNED);
1142 
1143 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1144 		aiop->aioprocflags |= AIOP_FREE;
1145 
1146 		/*
1147 		 * If daemon is inactive for a long time, allow it to exit,
1148 		 * thereby freeing resources.
1149 		 */
1150 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1151 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1152 		    (aiop->aioprocflags & AIOP_FREE) &&
1153 		    num_aio_procs > target_aio_procs)
1154 			break;
1155 	}
1156 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1157 	num_aio_procs--;
1158 	mtx_unlock(&aio_job_mtx);
1159 	uma_zfree(aiop_zone, aiop);
1160 	free_unr(aiod_unr, id);
1161 	vmspace_free(myvm);
1162 
1163 	KASSERT(p->p_vmspace == myvm,
1164 	    ("AIOD: bad vmspace for exiting daemon"));
1165 	KASSERT(myvm->vm_refcnt > 1,
1166 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1167 	kproc_exit(0);
1168 }
1169 
1170 /*
1171  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1172  * AIO daemon modifies its environment itself.
1173  */
1174 static int
1175 aio_newproc(int *start)
1176 {
1177 	int error;
1178 	struct proc *p;
1179 	int id;
1180 
1181 	id = alloc_unr(aiod_unr);
1182 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1183 		RFNOWAIT, 0, "aiod%d", id);
1184 	if (error == 0) {
1185 		/*
1186 		 * Wait until daemon is started.
1187 		 */
1188 		sema_wait(&aio_newproc_sem);
1189 		mtx_lock(&aio_job_mtx);
1190 		num_aio_procs++;
1191 		if (start != NULL)
1192 			(*start)--;
1193 		mtx_unlock(&aio_job_mtx);
1194 	} else {
1195 		free_unr(aiod_unr, id);
1196 	}
1197 	return (error);
1198 }
1199 
1200 /*
1201  * Try the high-performance, low-overhead physio method for eligible
1202  * VCHR devices.  This method doesn't use an aio helper thread, and
1203  * thus has very low overhead.
1204  *
1205  * Assumes that the caller, aio_aqueue(), has incremented the file
1206  * structure's reference count, preventing its deallocation for the
1207  * duration of this call.
1208  */
1209 static int
1210 aio_qphysio(struct proc *p, struct kaiocb *job)
1211 {
1212 	struct aiocb *cb;
1213 	struct file *fp;
1214 	struct bio *bp;
1215 	struct buf *pbuf;
1216 	struct vnode *vp;
1217 	struct cdevsw *csw;
1218 	struct cdev *dev;
1219 	struct kaioinfo *ki;
1220 	int error, ref, poff;
1221 	vm_prot_t prot;
1222 
1223 	cb = &job->uaiocb;
1224 	fp = job->fd_file;
1225 
1226 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1227 		return (-1);
1228 
1229 	vp = fp->f_vnode;
1230 	if (vp->v_type != VCHR)
1231 		return (-1);
1232 	if (vp->v_bufobj.bo_bsize == 0)
1233 		return (-1);
1234 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1235 		return (-1);
1236 
1237 	ref = 0;
1238 	csw = devvn_refthread(vp, &dev, &ref);
1239 	if (csw == NULL)
1240 		return (ENXIO);
1241 
1242 	if ((csw->d_flags & D_DISK) == 0) {
1243 		error = -1;
1244 		goto unref;
1245 	}
1246 	if (cb->aio_nbytes > dev->si_iosize_max) {
1247 		error = -1;
1248 		goto unref;
1249 	}
1250 
1251 	ki = p->p_aioinfo;
1252 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1253 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1254 		if (cb->aio_nbytes > MAXPHYS) {
1255 			error = -1;
1256 			goto unref;
1257 		}
1258 
1259 		pbuf = NULL;
1260 	} else {
1261 		if (cb->aio_nbytes > MAXPHYS - poff) {
1262 			error = -1;
1263 			goto unref;
1264 		}
1265 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1266 			error = -1;
1267 			goto unref;
1268 		}
1269 
1270 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1271 		BUF_KERNPROC(pbuf);
1272 		AIO_LOCK(ki);
1273 		ki->kaio_buffer_count++;
1274 		AIO_UNLOCK(ki);
1275 	}
1276 	job->bp = bp = g_alloc_bio();
1277 
1278 	bp->bio_length = cb->aio_nbytes;
1279 	bp->bio_bcount = cb->aio_nbytes;
1280 	bp->bio_done = aio_physwakeup;
1281 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1282 	bp->bio_offset = cb->aio_offset;
1283 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1284 	bp->bio_dev = dev;
1285 	bp->bio_caller1 = (void *)job;
1286 
1287 	prot = VM_PROT_READ;
1288 	if (cb->aio_lio_opcode == LIO_READ)
1289 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1290 	job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1291 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1292 	    nitems(job->pages));
1293 	if (job->npages < 0) {
1294 		error = EFAULT;
1295 		goto doerror;
1296 	}
1297 	if (pbuf != NULL) {
1298 		pmap_qenter((vm_offset_t)pbuf->b_data,
1299 		    job->pages, job->npages);
1300 		bp->bio_data = pbuf->b_data + poff;
1301 		atomic_add_int(&num_buf_aio, 1);
1302 	} else {
1303 		bp->bio_ma = job->pages;
1304 		bp->bio_ma_n = job->npages;
1305 		bp->bio_ma_offset = poff;
1306 		bp->bio_data = unmapped_buf;
1307 		bp->bio_flags |= BIO_UNMAPPED;
1308 	}
1309 
1310 	/* Perform transfer. */
1311 	csw->d_strategy(bp);
1312 	dev_relthread(dev, ref);
1313 	return (0);
1314 
1315 doerror:
1316 	if (pbuf != NULL) {
1317 		AIO_LOCK(ki);
1318 		ki->kaio_buffer_count--;
1319 		AIO_UNLOCK(ki);
1320 		relpbuf(pbuf, NULL);
1321 		job->pbuf = NULL;
1322 	}
1323 	g_destroy_bio(bp);
1324 	job->bp = NULL;
1325 unref:
1326 	dev_relthread(dev, ref);
1327 	return (error);
1328 }
1329 
1330 #ifdef COMPAT_FREEBSD6
1331 static int
1332 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1333 {
1334 
1335 	/*
1336 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1337 	 * supported by AIO with the old sigevent structure.
1338 	 */
1339 	nsig->sigev_notify = osig->sigev_notify;
1340 	switch (nsig->sigev_notify) {
1341 	case SIGEV_NONE:
1342 		break;
1343 	case SIGEV_SIGNAL:
1344 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1345 		break;
1346 	case SIGEV_KEVENT:
1347 		nsig->sigev_notify_kqueue =
1348 		    osig->__sigev_u.__sigev_notify_kqueue;
1349 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1350 		break;
1351 	default:
1352 		return (EINVAL);
1353 	}
1354 	return (0);
1355 }
1356 
1357 static int
1358 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1359 {
1360 	struct oaiocb *ojob;
1361 	int error;
1362 
1363 	bzero(kjob, sizeof(struct aiocb));
1364 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1365 	if (error)
1366 		return (error);
1367 	ojob = (struct oaiocb *)kjob;
1368 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1369 }
1370 #endif
1371 
1372 static int
1373 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1374 {
1375 
1376 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1377 }
1378 
1379 static long
1380 aiocb_fetch_status(struct aiocb *ujob)
1381 {
1382 
1383 	return (fuword(&ujob->_aiocb_private.status));
1384 }
1385 
1386 static long
1387 aiocb_fetch_error(struct aiocb *ujob)
1388 {
1389 
1390 	return (fuword(&ujob->_aiocb_private.error));
1391 }
1392 
1393 static int
1394 aiocb_store_status(struct aiocb *ujob, long status)
1395 {
1396 
1397 	return (suword(&ujob->_aiocb_private.status, status));
1398 }
1399 
1400 static int
1401 aiocb_store_error(struct aiocb *ujob, long error)
1402 {
1403 
1404 	return (suword(&ujob->_aiocb_private.error, error));
1405 }
1406 
1407 static int
1408 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1409 {
1410 
1411 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1412 }
1413 
1414 static int
1415 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1416 {
1417 
1418 	return (suword(ujobp, (long)ujob));
1419 }
1420 
1421 static struct aiocb_ops aiocb_ops = {
1422 	.copyin = aiocb_copyin,
1423 	.fetch_status = aiocb_fetch_status,
1424 	.fetch_error = aiocb_fetch_error,
1425 	.store_status = aiocb_store_status,
1426 	.store_error = aiocb_store_error,
1427 	.store_kernelinfo = aiocb_store_kernelinfo,
1428 	.store_aiocb = aiocb_store_aiocb,
1429 };
1430 
1431 #ifdef COMPAT_FREEBSD6
1432 static struct aiocb_ops aiocb_ops_osigevent = {
1433 	.copyin = aiocb_copyin_old_sigevent,
1434 	.fetch_status = aiocb_fetch_status,
1435 	.fetch_error = aiocb_fetch_error,
1436 	.store_status = aiocb_store_status,
1437 	.store_error = aiocb_store_error,
1438 	.store_kernelinfo = aiocb_store_kernelinfo,
1439 	.store_aiocb = aiocb_store_aiocb,
1440 };
1441 #endif
1442 
1443 /*
1444  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1445  * technique is done in this code.
1446  */
1447 int
1448 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1449     int type, struct aiocb_ops *ops)
1450 {
1451 	struct proc *p = td->td_proc;
1452 	cap_rights_t rights;
1453 	struct file *fp;
1454 	struct kaiocb *job;
1455 	struct kaioinfo *ki;
1456 	struct kevent kev;
1457 	int opcode;
1458 	int error;
1459 	int fd, kqfd;
1460 	int jid;
1461 	u_short evflags;
1462 
1463 	if (p->p_aioinfo == NULL)
1464 		aio_init_aioinfo(p);
1465 
1466 	ki = p->p_aioinfo;
1467 
1468 	ops->store_status(ujob, -1);
1469 	ops->store_error(ujob, 0);
1470 	ops->store_kernelinfo(ujob, -1);
1471 
1472 	if (num_queue_count >= max_queue_count ||
1473 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1474 		ops->store_error(ujob, EAGAIN);
1475 		return (EAGAIN);
1476 	}
1477 
1478 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1479 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1480 
1481 	error = ops->copyin(ujob, &job->uaiocb);
1482 	if (error) {
1483 		ops->store_error(ujob, error);
1484 		uma_zfree(aiocb_zone, job);
1485 		return (error);
1486 	}
1487 
1488 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1489 		uma_zfree(aiocb_zone, job);
1490 		return (EINVAL);
1491 	}
1492 
1493 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1494 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1495 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1496 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1497 		ops->store_error(ujob, EINVAL);
1498 		uma_zfree(aiocb_zone, job);
1499 		return (EINVAL);
1500 	}
1501 
1502 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1503 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1504 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1505 		uma_zfree(aiocb_zone, job);
1506 		return (EINVAL);
1507 	}
1508 
1509 	ksiginfo_init(&job->ksi);
1510 
1511 	/* Save userspace address of the job info. */
1512 	job->ujob = ujob;
1513 
1514 	/* Get the opcode. */
1515 	if (type != LIO_NOP)
1516 		job->uaiocb.aio_lio_opcode = type;
1517 	opcode = job->uaiocb.aio_lio_opcode;
1518 
1519 	/*
1520 	 * Validate the opcode and fetch the file object for the specified
1521 	 * file descriptor.
1522 	 *
1523 	 * XXXRW: Moved the opcode validation up here so that we don't
1524 	 * retrieve a file descriptor without knowing what the capabiltity
1525 	 * should be.
1526 	 */
1527 	fd = job->uaiocb.aio_fildes;
1528 	switch (opcode) {
1529 	case LIO_WRITE:
1530 		error = fget_write(td, fd,
1531 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1532 		break;
1533 	case LIO_READ:
1534 		error = fget_read(td, fd,
1535 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1536 		break;
1537 	case LIO_SYNC:
1538 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1539 		break;
1540 	case LIO_MLOCK:
1541 		fp = NULL;
1542 		break;
1543 	case LIO_NOP:
1544 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1545 		break;
1546 	default:
1547 		error = EINVAL;
1548 	}
1549 	if (error) {
1550 		uma_zfree(aiocb_zone, job);
1551 		ops->store_error(ujob, error);
1552 		return (error);
1553 	}
1554 
1555 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1556 		error = EINVAL;
1557 		goto aqueue_fail;
1558 	}
1559 
1560 	if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1561 	    job->uaiocb.aio_offset < 0 &&
1562 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1563 		error = EINVAL;
1564 		goto aqueue_fail;
1565 	}
1566 
1567 	job->fd_file = fp;
1568 
1569 	mtx_lock(&aio_job_mtx);
1570 	jid = jobrefid++;
1571 	job->seqno = jobseqno++;
1572 	mtx_unlock(&aio_job_mtx);
1573 	error = ops->store_kernelinfo(ujob, jid);
1574 	if (error) {
1575 		error = EINVAL;
1576 		goto aqueue_fail;
1577 	}
1578 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1579 
1580 	if (opcode == LIO_NOP) {
1581 		fdrop(fp, td);
1582 		uma_zfree(aiocb_zone, job);
1583 		return (0);
1584 	}
1585 
1586 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1587 		goto no_kqueue;
1588 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1589 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1590 		error = EINVAL;
1591 		goto aqueue_fail;
1592 	}
1593 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1594 	kev.ident = (uintptr_t)job->ujob;
1595 	kev.filter = EVFILT_AIO;
1596 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1597 	kev.data = (intptr_t)job;
1598 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1599 	error = kqfd_register(kqfd, &kev, td, 1);
1600 	if (error)
1601 		goto aqueue_fail;
1602 
1603 no_kqueue:
1604 
1605 	ops->store_error(ujob, EINPROGRESS);
1606 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1607 	job->userproc = p;
1608 	job->cred = crhold(td->td_ucred);
1609 	job->jobflags = KAIOCB_QUEUEING;
1610 	job->lio = lj;
1611 
1612 	if (opcode == LIO_MLOCK) {
1613 		aio_schedule(job, aio_process_mlock);
1614 		error = 0;
1615 	} else if (fp->f_ops->fo_aio_queue == NULL)
1616 		error = aio_queue_file(fp, job);
1617 	else
1618 		error = fo_aio_queue(fp, job);
1619 	if (error)
1620 		goto aqueue_fail;
1621 
1622 	AIO_LOCK(ki);
1623 	job->jobflags &= ~KAIOCB_QUEUEING;
1624 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1625 	ki->kaio_count++;
1626 	if (lj)
1627 		lj->lioj_count++;
1628 	atomic_add_int(&num_queue_count, 1);
1629 	if (job->jobflags & KAIOCB_FINISHED) {
1630 		/*
1631 		 * The queue callback completed the request synchronously.
1632 		 * The bulk of the completion is deferred in that case
1633 		 * until this point.
1634 		 */
1635 		aio_bio_done_notify(p, job);
1636 	} else
1637 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1638 	AIO_UNLOCK(ki);
1639 	return (0);
1640 
1641 aqueue_fail:
1642 	knlist_delete(&job->klist, curthread, 0);
1643 	if (fp)
1644 		fdrop(fp, td);
1645 	uma_zfree(aiocb_zone, job);
1646 	ops->store_error(ujob, error);
1647 	return (error);
1648 }
1649 
1650 static void
1651 aio_cancel_daemon_job(struct kaiocb *job)
1652 {
1653 
1654 	mtx_lock(&aio_job_mtx);
1655 	if (!aio_cancel_cleared(job))
1656 		TAILQ_REMOVE(&aio_jobs, job, list);
1657 	mtx_unlock(&aio_job_mtx);
1658 	aio_cancel(job);
1659 }
1660 
1661 void
1662 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1663 {
1664 
1665 	mtx_lock(&aio_job_mtx);
1666 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1667 		mtx_unlock(&aio_job_mtx);
1668 		aio_cancel(job);
1669 		return;
1670 	}
1671 	job->handle_fn = func;
1672 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1673 	aio_kick_nowait(job->userproc);
1674 	mtx_unlock(&aio_job_mtx);
1675 }
1676 
1677 static void
1678 aio_cancel_sync(struct kaiocb *job)
1679 {
1680 	struct kaioinfo *ki;
1681 
1682 	ki = job->userproc->p_aioinfo;
1683 	AIO_LOCK(ki);
1684 	if (!aio_cancel_cleared(job))
1685 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1686 	AIO_UNLOCK(ki);
1687 	aio_cancel(job);
1688 }
1689 
1690 int
1691 aio_queue_file(struct file *fp, struct kaiocb *job)
1692 {
1693 	struct kaioinfo *ki;
1694 	struct kaiocb *job2;
1695 	struct vnode *vp;
1696 	struct mount *mp;
1697 	int error, opcode;
1698 	bool safe;
1699 
1700 	ki = job->userproc->p_aioinfo;
1701 	opcode = job->uaiocb.aio_lio_opcode;
1702 	if (opcode == LIO_SYNC)
1703 		goto queueit;
1704 
1705 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1706 		goto done;
1707 #if 0
1708 	/*
1709 	 * XXX: This means qphysio() failed with EFAULT.  The current
1710 	 * behavior is to retry the operation via fo_read/fo_write.
1711 	 * Wouldn't it be better to just complete the request with an
1712 	 * error here?
1713 	 */
1714 	if (error > 0)
1715 		goto done;
1716 #endif
1717 queueit:
1718 	safe = false;
1719 	if (fp->f_type == DTYPE_VNODE) {
1720 		vp = fp->f_vnode;
1721 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1722 			mp = fp->f_vnode->v_mount;
1723 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1724 				safe = true;
1725 		}
1726 	}
1727 	if (!(safe || enable_aio_unsafe)) {
1728 		counted_warning(&unsafe_warningcnt,
1729 		    "is attempting to use unsafe AIO requests");
1730 		return (EOPNOTSUPP);
1731 	}
1732 
1733 	if (opcode == LIO_SYNC) {
1734 		AIO_LOCK(ki);
1735 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1736 			if (job2->fd_file == job->fd_file &&
1737 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1738 			    job2->seqno < job->seqno) {
1739 				job2->jobflags |= KAIOCB_CHECKSYNC;
1740 				job->pending++;
1741 			}
1742 		}
1743 		if (job->pending != 0) {
1744 			if (!aio_set_cancel_function_locked(job,
1745 				aio_cancel_sync)) {
1746 				AIO_UNLOCK(ki);
1747 				aio_cancel(job);
1748 				return (0);
1749 			}
1750 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1751 			AIO_UNLOCK(ki);
1752 			return (0);
1753 		}
1754 		AIO_UNLOCK(ki);
1755 	}
1756 
1757 	switch (opcode) {
1758 	case LIO_READ:
1759 	case LIO_WRITE:
1760 		aio_schedule(job, aio_process_rw);
1761 		error = 0;
1762 		break;
1763 	case LIO_SYNC:
1764 		aio_schedule(job, aio_process_sync);
1765 		error = 0;
1766 		break;
1767 	default:
1768 		error = EINVAL;
1769 	}
1770 done:
1771 	return (error);
1772 }
1773 
1774 static void
1775 aio_kick_nowait(struct proc *userp)
1776 {
1777 	struct kaioinfo *ki = userp->p_aioinfo;
1778 	struct aioproc *aiop;
1779 
1780 	mtx_assert(&aio_job_mtx, MA_OWNED);
1781 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1782 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1783 		aiop->aioprocflags &= ~AIOP_FREE;
1784 		wakeup(aiop->aioproc);
1785 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1786 	    ki->kaio_active_count + num_aio_resv_start <
1787 	    ki->kaio_maxactive_count) {
1788 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1789 	}
1790 }
1791 
1792 static int
1793 aio_kick(struct proc *userp)
1794 {
1795 	struct kaioinfo *ki = userp->p_aioinfo;
1796 	struct aioproc *aiop;
1797 	int error, ret = 0;
1798 
1799 	mtx_assert(&aio_job_mtx, MA_OWNED);
1800 retryproc:
1801 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1802 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1803 		aiop->aioprocflags &= ~AIOP_FREE;
1804 		wakeup(aiop->aioproc);
1805 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1806 	    ki->kaio_active_count + num_aio_resv_start <
1807 	    ki->kaio_maxactive_count) {
1808 		num_aio_resv_start++;
1809 		mtx_unlock(&aio_job_mtx);
1810 		error = aio_newproc(&num_aio_resv_start);
1811 		mtx_lock(&aio_job_mtx);
1812 		if (error) {
1813 			num_aio_resv_start--;
1814 			goto retryproc;
1815 		}
1816 	} else {
1817 		ret = -1;
1818 	}
1819 	return (ret);
1820 }
1821 
1822 static void
1823 aio_kick_helper(void *context, int pending)
1824 {
1825 	struct proc *userp = context;
1826 
1827 	mtx_lock(&aio_job_mtx);
1828 	while (--pending >= 0) {
1829 		if (aio_kick(userp))
1830 			break;
1831 	}
1832 	mtx_unlock(&aio_job_mtx);
1833 }
1834 
1835 /*
1836  * Support the aio_return system call, as a side-effect, kernel resources are
1837  * released.
1838  */
1839 static int
1840 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1841 {
1842 	struct proc *p = td->td_proc;
1843 	struct kaiocb *job;
1844 	struct kaioinfo *ki;
1845 	long status, error;
1846 
1847 	ki = p->p_aioinfo;
1848 	if (ki == NULL)
1849 		return (EINVAL);
1850 	AIO_LOCK(ki);
1851 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1852 		if (job->ujob == ujob)
1853 			break;
1854 	}
1855 	if (job != NULL) {
1856 		MPASS(job->jobflags & KAIOCB_FINISHED);
1857 		status = job->uaiocb._aiocb_private.status;
1858 		error = job->uaiocb._aiocb_private.error;
1859 		td->td_retval[0] = status;
1860 		td->td_ru.ru_oublock += job->outblock;
1861 		td->td_ru.ru_inblock += job->inblock;
1862 		td->td_ru.ru_msgsnd += job->msgsnd;
1863 		td->td_ru.ru_msgrcv += job->msgrcv;
1864 		aio_free_entry(job);
1865 		AIO_UNLOCK(ki);
1866 		ops->store_error(ujob, error);
1867 		ops->store_status(ujob, status);
1868 	} else {
1869 		error = EINVAL;
1870 		AIO_UNLOCK(ki);
1871 	}
1872 	return (error);
1873 }
1874 
1875 int
1876 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1877 {
1878 
1879 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1880 }
1881 
1882 /*
1883  * Allow a process to wakeup when any of the I/O requests are completed.
1884  */
1885 static int
1886 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1887     struct timespec *ts)
1888 {
1889 	struct proc *p = td->td_proc;
1890 	struct timeval atv;
1891 	struct kaioinfo *ki;
1892 	struct kaiocb *firstjob, *job;
1893 	int error, i, timo;
1894 
1895 	timo = 0;
1896 	if (ts) {
1897 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1898 			return (EINVAL);
1899 
1900 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1901 		if (itimerfix(&atv))
1902 			return (EINVAL);
1903 		timo = tvtohz(&atv);
1904 	}
1905 
1906 	ki = p->p_aioinfo;
1907 	if (ki == NULL)
1908 		return (EAGAIN);
1909 
1910 	if (njoblist == 0)
1911 		return (0);
1912 
1913 	AIO_LOCK(ki);
1914 	for (;;) {
1915 		firstjob = NULL;
1916 		error = 0;
1917 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1918 			for (i = 0; i < njoblist; i++) {
1919 				if (job->ujob == ujoblist[i]) {
1920 					if (firstjob == NULL)
1921 						firstjob = job;
1922 					if (job->jobflags & KAIOCB_FINISHED)
1923 						goto RETURN;
1924 				}
1925 			}
1926 		}
1927 		/* All tasks were finished. */
1928 		if (firstjob == NULL)
1929 			break;
1930 
1931 		ki->kaio_flags |= KAIO_WAKEUP;
1932 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1933 		    "aiospn", timo);
1934 		if (error == ERESTART)
1935 			error = EINTR;
1936 		if (error)
1937 			break;
1938 	}
1939 RETURN:
1940 	AIO_UNLOCK(ki);
1941 	return (error);
1942 }
1943 
1944 int
1945 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1946 {
1947 	struct timespec ts, *tsp;
1948 	struct aiocb **ujoblist;
1949 	int error;
1950 
1951 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1952 		return (EINVAL);
1953 
1954 	if (uap->timeout) {
1955 		/* Get timespec struct. */
1956 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1957 			return (error);
1958 		tsp = &ts;
1959 	} else
1960 		tsp = NULL;
1961 
1962 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1963 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1964 	if (error == 0)
1965 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1966 	free(ujoblist, M_AIOS);
1967 	return (error);
1968 }
1969 
1970 /*
1971  * aio_cancel cancels any non-physio aio operations not currently in
1972  * progress.
1973  */
1974 int
1975 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1976 {
1977 	struct proc *p = td->td_proc;
1978 	struct kaioinfo *ki;
1979 	struct kaiocb *job, *jobn;
1980 	struct file *fp;
1981 	cap_rights_t rights;
1982 	int error;
1983 	int cancelled = 0;
1984 	int notcancelled = 0;
1985 	struct vnode *vp;
1986 
1987 	/* Lookup file object. */
1988 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1989 	if (error)
1990 		return (error);
1991 
1992 	ki = p->p_aioinfo;
1993 	if (ki == NULL)
1994 		goto done;
1995 
1996 	if (fp->f_type == DTYPE_VNODE) {
1997 		vp = fp->f_vnode;
1998 		if (vn_isdisk(vp, &error)) {
1999 			fdrop(fp, td);
2000 			td->td_retval[0] = AIO_NOTCANCELED;
2001 			return (0);
2002 		}
2003 	}
2004 
2005 	AIO_LOCK(ki);
2006 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2007 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2008 		    ((uap->aiocbp == NULL) ||
2009 		     (uap->aiocbp == job->ujob))) {
2010 			if (aio_cancel_job(p, ki, job)) {
2011 				cancelled++;
2012 			} else {
2013 				notcancelled++;
2014 			}
2015 			if (uap->aiocbp != NULL)
2016 				break;
2017 		}
2018 	}
2019 	AIO_UNLOCK(ki);
2020 
2021 done:
2022 	fdrop(fp, td);
2023 
2024 	if (uap->aiocbp != NULL) {
2025 		if (cancelled) {
2026 			td->td_retval[0] = AIO_CANCELED;
2027 			return (0);
2028 		}
2029 	}
2030 
2031 	if (notcancelled) {
2032 		td->td_retval[0] = AIO_NOTCANCELED;
2033 		return (0);
2034 	}
2035 
2036 	if (cancelled) {
2037 		td->td_retval[0] = AIO_CANCELED;
2038 		return (0);
2039 	}
2040 
2041 	td->td_retval[0] = AIO_ALLDONE;
2042 
2043 	return (0);
2044 }
2045 
2046 /*
2047  * aio_error is implemented in the kernel level for compatibility purposes
2048  * only.  For a user mode async implementation, it would be best to do it in
2049  * a userland subroutine.
2050  */
2051 static int
2052 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2053 {
2054 	struct proc *p = td->td_proc;
2055 	struct kaiocb *job;
2056 	struct kaioinfo *ki;
2057 	int status;
2058 
2059 	ki = p->p_aioinfo;
2060 	if (ki == NULL) {
2061 		td->td_retval[0] = EINVAL;
2062 		return (0);
2063 	}
2064 
2065 	AIO_LOCK(ki);
2066 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2067 		if (job->ujob == ujob) {
2068 			if (job->jobflags & KAIOCB_FINISHED)
2069 				td->td_retval[0] =
2070 					job->uaiocb._aiocb_private.error;
2071 			else
2072 				td->td_retval[0] = EINPROGRESS;
2073 			AIO_UNLOCK(ki);
2074 			return (0);
2075 		}
2076 	}
2077 	AIO_UNLOCK(ki);
2078 
2079 	/*
2080 	 * Hack for failure of aio_aqueue.
2081 	 */
2082 	status = ops->fetch_status(ujob);
2083 	if (status == -1) {
2084 		td->td_retval[0] = ops->fetch_error(ujob);
2085 		return (0);
2086 	}
2087 
2088 	td->td_retval[0] = EINVAL;
2089 	return (0);
2090 }
2091 
2092 int
2093 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2094 {
2095 
2096 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2097 }
2098 
2099 /* syscall - asynchronous read from a file (REALTIME) */
2100 #ifdef COMPAT_FREEBSD6
2101 int
2102 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2103 {
2104 
2105 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2106 	    &aiocb_ops_osigevent));
2107 }
2108 #endif
2109 
2110 int
2111 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2112 {
2113 
2114 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2115 }
2116 
2117 /* syscall - asynchronous write to a file (REALTIME) */
2118 #ifdef COMPAT_FREEBSD6
2119 int
2120 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2121 {
2122 
2123 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2124 	    &aiocb_ops_osigevent));
2125 }
2126 #endif
2127 
2128 int
2129 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2130 {
2131 
2132 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2133 }
2134 
2135 int
2136 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2137 {
2138 
2139 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2140 }
2141 
2142 static int
2143 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2144     struct aiocb **acb_list, int nent, struct sigevent *sig,
2145     struct aiocb_ops *ops)
2146 {
2147 	struct proc *p = td->td_proc;
2148 	struct aiocb *job;
2149 	struct kaioinfo *ki;
2150 	struct aioliojob *lj;
2151 	struct kevent kev;
2152 	int error;
2153 	int nerror;
2154 	int i;
2155 
2156 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2157 		return (EINVAL);
2158 
2159 	if (nent < 0 || nent > max_aio_queue_per_proc)
2160 		return (EINVAL);
2161 
2162 	if (p->p_aioinfo == NULL)
2163 		aio_init_aioinfo(p);
2164 
2165 	ki = p->p_aioinfo;
2166 
2167 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2168 	lj->lioj_flags = 0;
2169 	lj->lioj_count = 0;
2170 	lj->lioj_finished_count = 0;
2171 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2172 	ksiginfo_init(&lj->lioj_ksi);
2173 
2174 	/*
2175 	 * Setup signal.
2176 	 */
2177 	if (sig && (mode == LIO_NOWAIT)) {
2178 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2179 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2180 			/* Assume only new style KEVENT */
2181 			kev.filter = EVFILT_LIO;
2182 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2183 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2184 			kev.data = (intptr_t)lj;
2185 			/* pass user defined sigval data */
2186 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2187 			error = kqfd_register(
2188 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2189 			if (error) {
2190 				uma_zfree(aiolio_zone, lj);
2191 				return (error);
2192 			}
2193 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2194 			;
2195 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2196 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2197 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2198 					uma_zfree(aiolio_zone, lj);
2199 					return EINVAL;
2200 				}
2201 				lj->lioj_flags |= LIOJ_SIGNAL;
2202 		} else {
2203 			uma_zfree(aiolio_zone, lj);
2204 			return EINVAL;
2205 		}
2206 	}
2207 
2208 	AIO_LOCK(ki);
2209 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2210 	/*
2211 	 * Add extra aiocb count to avoid the lio to be freed
2212 	 * by other threads doing aio_waitcomplete or aio_return,
2213 	 * and prevent event from being sent until we have queued
2214 	 * all tasks.
2215 	 */
2216 	lj->lioj_count = 1;
2217 	AIO_UNLOCK(ki);
2218 
2219 	/*
2220 	 * Get pointers to the list of I/O requests.
2221 	 */
2222 	nerror = 0;
2223 	for (i = 0; i < nent; i++) {
2224 		job = acb_list[i];
2225 		if (job != NULL) {
2226 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2227 			if (error != 0)
2228 				nerror++;
2229 		}
2230 	}
2231 
2232 	error = 0;
2233 	AIO_LOCK(ki);
2234 	if (mode == LIO_WAIT) {
2235 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2236 			ki->kaio_flags |= KAIO_WAKEUP;
2237 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2238 			    PRIBIO | PCATCH, "aiospn", 0);
2239 			if (error == ERESTART)
2240 				error = EINTR;
2241 			if (error)
2242 				break;
2243 		}
2244 	} else {
2245 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2246 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2247 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2248 				KNOTE_LOCKED(&lj->klist, 1);
2249 			}
2250 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2251 			    == LIOJ_SIGNAL
2252 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2253 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2254 				aio_sendsig(p, &lj->lioj_signal,
2255 					    &lj->lioj_ksi);
2256 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2257 			}
2258 		}
2259 	}
2260 	lj->lioj_count--;
2261 	if (lj->lioj_count == 0) {
2262 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2263 		knlist_delete(&lj->klist, curthread, 1);
2264 		PROC_LOCK(p);
2265 		sigqueue_take(&lj->lioj_ksi);
2266 		PROC_UNLOCK(p);
2267 		AIO_UNLOCK(ki);
2268 		uma_zfree(aiolio_zone, lj);
2269 	} else
2270 		AIO_UNLOCK(ki);
2271 
2272 	if (nerror)
2273 		return (EIO);
2274 	return (error);
2275 }
2276 
2277 /* syscall - list directed I/O (REALTIME) */
2278 #ifdef COMPAT_FREEBSD6
2279 int
2280 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2281 {
2282 	struct aiocb **acb_list;
2283 	struct sigevent *sigp, sig;
2284 	struct osigevent osig;
2285 	int error, nent;
2286 
2287 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2288 		return (EINVAL);
2289 
2290 	nent = uap->nent;
2291 	if (nent < 0 || nent > max_aio_queue_per_proc)
2292 		return (EINVAL);
2293 
2294 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2295 		error = copyin(uap->sig, &osig, sizeof(osig));
2296 		if (error)
2297 			return (error);
2298 		error = convert_old_sigevent(&osig, &sig);
2299 		if (error)
2300 			return (error);
2301 		sigp = &sig;
2302 	} else
2303 		sigp = NULL;
2304 
2305 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2306 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2307 	if (error == 0)
2308 		error = kern_lio_listio(td, uap->mode,
2309 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2310 		    &aiocb_ops_osigevent);
2311 	free(acb_list, M_LIO);
2312 	return (error);
2313 }
2314 #endif
2315 
2316 /* syscall - list directed I/O (REALTIME) */
2317 int
2318 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2319 {
2320 	struct aiocb **acb_list;
2321 	struct sigevent *sigp, sig;
2322 	int error, nent;
2323 
2324 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2325 		return (EINVAL);
2326 
2327 	nent = uap->nent;
2328 	if (nent < 0 || nent > max_aio_queue_per_proc)
2329 		return (EINVAL);
2330 
2331 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2332 		error = copyin(uap->sig, &sig, sizeof(sig));
2333 		if (error)
2334 			return (error);
2335 		sigp = &sig;
2336 	} else
2337 		sigp = NULL;
2338 
2339 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2340 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2341 	if (error == 0)
2342 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2343 		    nent, sigp, &aiocb_ops);
2344 	free(acb_list, M_LIO);
2345 	return (error);
2346 }
2347 
2348 static void
2349 aio_physwakeup(struct bio *bp)
2350 {
2351 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2352 	struct proc *userp;
2353 	struct kaioinfo *ki;
2354 	size_t nbytes;
2355 	int error, nblks;
2356 
2357 	/* Release mapping into kernel space. */
2358 	userp = job->userproc;
2359 	ki = userp->p_aioinfo;
2360 	if (job->pbuf) {
2361 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2362 		relpbuf(job->pbuf, NULL);
2363 		job->pbuf = NULL;
2364 		atomic_subtract_int(&num_buf_aio, 1);
2365 		AIO_LOCK(ki);
2366 		ki->kaio_buffer_count--;
2367 		AIO_UNLOCK(ki);
2368 	}
2369 	vm_page_unhold_pages(job->pages, job->npages);
2370 
2371 	bp = job->bp;
2372 	job->bp = NULL;
2373 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2374 	error = 0;
2375 	if (bp->bio_flags & BIO_ERROR)
2376 		error = bp->bio_error;
2377 	nblks = btodb(nbytes);
2378 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2379 		job->outblock += nblks;
2380 	else
2381 		job->inblock += nblks;
2382 
2383 	if (error)
2384 		aio_complete(job, -1, error);
2385 	else
2386 		aio_complete(job, nbytes, 0);
2387 
2388 	g_destroy_bio(bp);
2389 }
2390 
2391 /* syscall - wait for the next completion of an aio request */
2392 static int
2393 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2394     struct timespec *ts, struct aiocb_ops *ops)
2395 {
2396 	struct proc *p = td->td_proc;
2397 	struct timeval atv;
2398 	struct kaioinfo *ki;
2399 	struct kaiocb *job;
2400 	struct aiocb *ujob;
2401 	long error, status;
2402 	int timo;
2403 
2404 	ops->store_aiocb(ujobp, NULL);
2405 
2406 	if (ts == NULL) {
2407 		timo = 0;
2408 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2409 		timo = -1;
2410 	} else {
2411 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2412 			return (EINVAL);
2413 
2414 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2415 		if (itimerfix(&atv))
2416 			return (EINVAL);
2417 		timo = tvtohz(&atv);
2418 	}
2419 
2420 	if (p->p_aioinfo == NULL)
2421 		aio_init_aioinfo(p);
2422 	ki = p->p_aioinfo;
2423 
2424 	error = 0;
2425 	job = NULL;
2426 	AIO_LOCK(ki);
2427 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2428 		if (timo == -1) {
2429 			error = EWOULDBLOCK;
2430 			break;
2431 		}
2432 		ki->kaio_flags |= KAIO_WAKEUP;
2433 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2434 		    "aiowc", timo);
2435 		if (timo && error == ERESTART)
2436 			error = EINTR;
2437 		if (error)
2438 			break;
2439 	}
2440 
2441 	if (job != NULL) {
2442 		MPASS(job->jobflags & KAIOCB_FINISHED);
2443 		ujob = job->ujob;
2444 		status = job->uaiocb._aiocb_private.status;
2445 		error = job->uaiocb._aiocb_private.error;
2446 		td->td_retval[0] = status;
2447 		td->td_ru.ru_oublock += job->outblock;
2448 		td->td_ru.ru_inblock += job->inblock;
2449 		td->td_ru.ru_msgsnd += job->msgsnd;
2450 		td->td_ru.ru_msgrcv += job->msgrcv;
2451 		aio_free_entry(job);
2452 		AIO_UNLOCK(ki);
2453 		ops->store_aiocb(ujobp, ujob);
2454 		ops->store_error(ujob, error);
2455 		ops->store_status(ujob, status);
2456 	} else
2457 		AIO_UNLOCK(ki);
2458 
2459 	return (error);
2460 }
2461 
2462 int
2463 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2464 {
2465 	struct timespec ts, *tsp;
2466 	int error;
2467 
2468 	if (uap->timeout) {
2469 		/* Get timespec struct. */
2470 		error = copyin(uap->timeout, &ts, sizeof(ts));
2471 		if (error)
2472 			return (error);
2473 		tsp = &ts;
2474 	} else
2475 		tsp = NULL;
2476 
2477 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2478 }
2479 
2480 static int
2481 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2482     struct aiocb_ops *ops)
2483 {
2484 
2485 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2486 		return (EINVAL);
2487 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2488 }
2489 
2490 int
2491 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2492 {
2493 
2494 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2495 }
2496 
2497 /* kqueue attach function */
2498 static int
2499 filt_aioattach(struct knote *kn)
2500 {
2501 	struct kaiocb *job;
2502 
2503 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2504 
2505 	/*
2506 	 * The job pointer must be validated before using it, so
2507 	 * registration is restricted to the kernel; the user cannot
2508 	 * set EV_FLAG1.
2509 	 */
2510 	if ((kn->kn_flags & EV_FLAG1) == 0)
2511 		return (EPERM);
2512 	kn->kn_ptr.p_aio = job;
2513 	kn->kn_flags &= ~EV_FLAG1;
2514 
2515 	knlist_add(&job->klist, kn, 0);
2516 
2517 	return (0);
2518 }
2519 
2520 /* kqueue detach function */
2521 static void
2522 filt_aiodetach(struct knote *kn)
2523 {
2524 	struct knlist *knl;
2525 
2526 	knl = &kn->kn_ptr.p_aio->klist;
2527 	knl->kl_lock(knl->kl_lockarg);
2528 	if (!knlist_empty(knl))
2529 		knlist_remove(knl, kn, 1);
2530 	knl->kl_unlock(knl->kl_lockarg);
2531 }
2532 
2533 /* kqueue filter function */
2534 /*ARGSUSED*/
2535 static int
2536 filt_aio(struct knote *kn, long hint)
2537 {
2538 	struct kaiocb *job = kn->kn_ptr.p_aio;
2539 
2540 	kn->kn_data = job->uaiocb._aiocb_private.error;
2541 	if (!(job->jobflags & KAIOCB_FINISHED))
2542 		return (0);
2543 	kn->kn_flags |= EV_EOF;
2544 	return (1);
2545 }
2546 
2547 /* kqueue attach function */
2548 static int
2549 filt_lioattach(struct knote *kn)
2550 {
2551 	struct aioliojob *lj;
2552 
2553 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2554 
2555 	/*
2556 	 * The aioliojob pointer must be validated before using it, so
2557 	 * registration is restricted to the kernel; the user cannot
2558 	 * set EV_FLAG1.
2559 	 */
2560 	if ((kn->kn_flags & EV_FLAG1) == 0)
2561 		return (EPERM);
2562 	kn->kn_ptr.p_lio = lj;
2563 	kn->kn_flags &= ~EV_FLAG1;
2564 
2565 	knlist_add(&lj->klist, kn, 0);
2566 
2567 	return (0);
2568 }
2569 
2570 /* kqueue detach function */
2571 static void
2572 filt_liodetach(struct knote *kn)
2573 {
2574 	struct knlist *knl;
2575 
2576 	knl = &kn->kn_ptr.p_lio->klist;
2577 	knl->kl_lock(knl->kl_lockarg);
2578 	if (!knlist_empty(knl))
2579 		knlist_remove(knl, kn, 1);
2580 	knl->kl_unlock(knl->kl_lockarg);
2581 }
2582 
2583 /* kqueue filter function */
2584 /*ARGSUSED*/
2585 static int
2586 filt_lio(struct knote *kn, long hint)
2587 {
2588 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2589 
2590 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2591 }
2592 
2593 #ifdef COMPAT_FREEBSD32
2594 #include <sys/mount.h>
2595 #include <sys/socket.h>
2596 #include <compat/freebsd32/freebsd32.h>
2597 #include <compat/freebsd32/freebsd32_proto.h>
2598 #include <compat/freebsd32/freebsd32_signal.h>
2599 #include <compat/freebsd32/freebsd32_syscall.h>
2600 #include <compat/freebsd32/freebsd32_util.h>
2601 
2602 struct __aiocb_private32 {
2603 	int32_t	status;
2604 	int32_t	error;
2605 	uint32_t kernelinfo;
2606 };
2607 
2608 #ifdef COMPAT_FREEBSD6
2609 typedef struct oaiocb32 {
2610 	int	aio_fildes;		/* File descriptor */
2611 	uint64_t aio_offset __packed;	/* File offset for I/O */
2612 	uint32_t aio_buf;		/* I/O buffer in process space */
2613 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2614 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2615 	int	aio_lio_opcode;		/* LIO opcode */
2616 	int	aio_reqprio;		/* Request priority -- ignored */
2617 	struct	__aiocb_private32 _aiocb_private;
2618 } oaiocb32_t;
2619 #endif
2620 
2621 typedef struct aiocb32 {
2622 	int32_t	aio_fildes;		/* File descriptor */
2623 	uint64_t aio_offset __packed;	/* File offset for I/O */
2624 	uint32_t aio_buf;		/* I/O buffer in process space */
2625 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2626 	int	__spare__[2];
2627 	uint32_t __spare2__;
2628 	int	aio_lio_opcode;		/* LIO opcode */
2629 	int	aio_reqprio;		/* Request priority -- ignored */
2630 	struct	__aiocb_private32 _aiocb_private;
2631 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2632 } aiocb32_t;
2633 
2634 #ifdef COMPAT_FREEBSD6
2635 static int
2636 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2637 {
2638 
2639 	/*
2640 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2641 	 * supported by AIO with the old sigevent structure.
2642 	 */
2643 	CP(*osig, *nsig, sigev_notify);
2644 	switch (nsig->sigev_notify) {
2645 	case SIGEV_NONE:
2646 		break;
2647 	case SIGEV_SIGNAL:
2648 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2649 		break;
2650 	case SIGEV_KEVENT:
2651 		nsig->sigev_notify_kqueue =
2652 		    osig->__sigev_u.__sigev_notify_kqueue;
2653 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2654 		break;
2655 	default:
2656 		return (EINVAL);
2657 	}
2658 	return (0);
2659 }
2660 
2661 static int
2662 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2663 {
2664 	struct oaiocb32 job32;
2665 	int error;
2666 
2667 	bzero(kjob, sizeof(struct aiocb));
2668 	error = copyin(ujob, &job32, sizeof(job32));
2669 	if (error)
2670 		return (error);
2671 
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_old_sigevent32(&job32.aio_sigevent,
2682 	    &kjob->aio_sigevent));
2683 }
2684 #endif
2685 
2686 static int
2687 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2688 {
2689 	struct aiocb32 job32;
2690 	int error;
2691 
2692 	error = copyin(ujob, &job32, sizeof(job32));
2693 	if (error)
2694 		return (error);
2695 	CP(job32, *kjob, aio_fildes);
2696 	CP(job32, *kjob, aio_offset);
2697 	PTRIN_CP(job32, *kjob, aio_buf);
2698 	CP(job32, *kjob, aio_nbytes);
2699 	CP(job32, *kjob, aio_lio_opcode);
2700 	CP(job32, *kjob, aio_reqprio);
2701 	CP(job32, *kjob, _aiocb_private.status);
2702 	CP(job32, *kjob, _aiocb_private.error);
2703 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2704 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2705 }
2706 
2707 static long
2708 aiocb32_fetch_status(struct aiocb *ujob)
2709 {
2710 	struct aiocb32 *ujob32;
2711 
2712 	ujob32 = (struct aiocb32 *)ujob;
2713 	return (fuword32(&ujob32->_aiocb_private.status));
2714 }
2715 
2716 static long
2717 aiocb32_fetch_error(struct aiocb *ujob)
2718 {
2719 	struct aiocb32 *ujob32;
2720 
2721 	ujob32 = (struct aiocb32 *)ujob;
2722 	return (fuword32(&ujob32->_aiocb_private.error));
2723 }
2724 
2725 static int
2726 aiocb32_store_status(struct aiocb *ujob, long status)
2727 {
2728 	struct aiocb32 *ujob32;
2729 
2730 	ujob32 = (struct aiocb32 *)ujob;
2731 	return (suword32(&ujob32->_aiocb_private.status, status));
2732 }
2733 
2734 static int
2735 aiocb32_store_error(struct aiocb *ujob, long error)
2736 {
2737 	struct aiocb32 *ujob32;
2738 
2739 	ujob32 = (struct aiocb32 *)ujob;
2740 	return (suword32(&ujob32->_aiocb_private.error, error));
2741 }
2742 
2743 static int
2744 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2745 {
2746 	struct aiocb32 *ujob32;
2747 
2748 	ujob32 = (struct aiocb32 *)ujob;
2749 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2750 }
2751 
2752 static int
2753 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2754 {
2755 
2756 	return (suword32(ujobp, (long)ujob));
2757 }
2758 
2759 static struct aiocb_ops aiocb32_ops = {
2760 	.copyin = aiocb32_copyin,
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 
2769 #ifdef COMPAT_FREEBSD6
2770 static struct aiocb_ops aiocb32_ops_osigevent = {
2771 	.copyin = aiocb32_copyin_old_sigevent,
2772 	.fetch_status = aiocb32_fetch_status,
2773 	.fetch_error = aiocb32_fetch_error,
2774 	.store_status = aiocb32_store_status,
2775 	.store_error = aiocb32_store_error,
2776 	.store_kernelinfo = aiocb32_store_kernelinfo,
2777 	.store_aiocb = aiocb32_store_aiocb,
2778 };
2779 #endif
2780 
2781 int
2782 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2783 {
2784 
2785 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2786 }
2787 
2788 int
2789 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2790 {
2791 	struct timespec32 ts32;
2792 	struct timespec ts, *tsp;
2793 	struct aiocb **ujoblist;
2794 	uint32_t *ujoblist32;
2795 	int error, i;
2796 
2797 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2798 		return (EINVAL);
2799 
2800 	if (uap->timeout) {
2801 		/* Get timespec struct. */
2802 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2803 			return (error);
2804 		CP(ts32, ts, tv_sec);
2805 		CP(ts32, ts, tv_nsec);
2806 		tsp = &ts;
2807 	} else
2808 		tsp = NULL;
2809 
2810 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2811 	ujoblist32 = (uint32_t *)ujoblist;
2812 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2813 	    sizeof(ujoblist32[0]));
2814 	if (error == 0) {
2815 		for (i = uap->nent - 1; i >= 0; i--)
2816 			ujoblist[i] = PTRIN(ujoblist32[i]);
2817 
2818 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2819 	}
2820 	free(ujoblist, M_AIOS);
2821 	return (error);
2822 }
2823 
2824 int
2825 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2826 {
2827 
2828 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2829 }
2830 
2831 #ifdef COMPAT_FREEBSD6
2832 int
2833 freebsd6_freebsd32_aio_read(struct thread *td,
2834     struct freebsd6_freebsd32_aio_read_args *uap)
2835 {
2836 
2837 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2838 	    &aiocb32_ops_osigevent));
2839 }
2840 #endif
2841 
2842 int
2843 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2844 {
2845 
2846 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2847 	    &aiocb32_ops));
2848 }
2849 
2850 #ifdef COMPAT_FREEBSD6
2851 int
2852 freebsd6_freebsd32_aio_write(struct thread *td,
2853     struct freebsd6_freebsd32_aio_write_args *uap)
2854 {
2855 
2856 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2857 	    &aiocb32_ops_osigevent));
2858 }
2859 #endif
2860 
2861 int
2862 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2863 {
2864 
2865 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2866 	    &aiocb32_ops));
2867 }
2868 
2869 int
2870 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2871 {
2872 
2873 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2874 	    &aiocb32_ops));
2875 }
2876 
2877 int
2878 freebsd32_aio_waitcomplete(struct thread *td,
2879     struct freebsd32_aio_waitcomplete_args *uap)
2880 {
2881 	struct timespec32 ts32;
2882 	struct timespec ts, *tsp;
2883 	int error;
2884 
2885 	if (uap->timeout) {
2886 		/* Get timespec struct. */
2887 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2888 		if (error)
2889 			return (error);
2890 		CP(ts32, ts, tv_sec);
2891 		CP(ts32, ts, tv_nsec);
2892 		tsp = &ts;
2893 	} else
2894 		tsp = NULL;
2895 
2896 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2897 	    &aiocb32_ops));
2898 }
2899 
2900 int
2901 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2902 {
2903 
2904 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2905 	    &aiocb32_ops));
2906 }
2907 
2908 #ifdef COMPAT_FREEBSD6
2909 int
2910 freebsd6_freebsd32_lio_listio(struct thread *td,
2911     struct freebsd6_freebsd32_lio_listio_args *uap)
2912 {
2913 	struct aiocb **acb_list;
2914 	struct sigevent *sigp, sig;
2915 	struct osigevent32 osig;
2916 	uint32_t *acb_list32;
2917 	int error, i, nent;
2918 
2919 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2920 		return (EINVAL);
2921 
2922 	nent = uap->nent;
2923 	if (nent < 0 || nent > max_aio_queue_per_proc)
2924 		return (EINVAL);
2925 
2926 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2927 		error = copyin(uap->sig, &osig, sizeof(osig));
2928 		if (error)
2929 			return (error);
2930 		error = convert_old_sigevent32(&osig, &sig);
2931 		if (error)
2932 			return (error);
2933 		sigp = &sig;
2934 	} else
2935 		sigp = NULL;
2936 
2937 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2938 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2939 	if (error) {
2940 		free(acb_list32, M_LIO);
2941 		return (error);
2942 	}
2943 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2944 	for (i = 0; i < nent; i++)
2945 		acb_list[i] = PTRIN(acb_list32[i]);
2946 	free(acb_list32, M_LIO);
2947 
2948 	error = kern_lio_listio(td, uap->mode,
2949 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2950 	    &aiocb32_ops_osigevent);
2951 	free(acb_list, M_LIO);
2952 	return (error);
2953 }
2954 #endif
2955 
2956 int
2957 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2958 {
2959 	struct aiocb **acb_list;
2960 	struct sigevent *sigp, sig;
2961 	struct sigevent32 sig32;
2962 	uint32_t *acb_list32;
2963 	int error, i, nent;
2964 
2965 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2966 		return (EINVAL);
2967 
2968 	nent = uap->nent;
2969 	if (nent < 0 || nent > max_aio_queue_per_proc)
2970 		return (EINVAL);
2971 
2972 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2973 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2974 		if (error)
2975 			return (error);
2976 		error = convert_sigevent32(&sig32, &sig);
2977 		if (error)
2978 			return (error);
2979 		sigp = &sig;
2980 	} else
2981 		sigp = NULL;
2982 
2983 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2984 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2985 	if (error) {
2986 		free(acb_list32, M_LIO);
2987 		return (error);
2988 	}
2989 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2990 	for (i = 0; i < nent; i++)
2991 		acb_list[i] = PTRIN(acb_list32[i]);
2992 	free(acb_list32, M_LIO);
2993 
2994 	error = kern_lio_listio(td, uap->mode,
2995 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2996 	    &aiocb32_ops);
2997 	free(acb_list, M_LIO);
2998 	return (error);
2999 }
3000 
3001 #endif
3002