xref: /freebsd/sys/kern/vfs_aio.c (revision 792bbaba989533a1fc93823df1720c8c4aaf0442)
1 /*-
2  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
3  *
4  * Redistribution and use in source and binary forms, with or without
5  * modification, are permitted provided that the following conditions
6  * are met:
7  * 1. Redistributions of source code must retain the above copyright
8  *    notice, this list of conditions and the following disclaimer.
9  * 2. John S. Dyson's name may not be used to endorse or promote products
10  *    derived from this software without specific prior written permission.
11  *
12  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
13  * bad that happens because of using this software isn't the responsibility
14  * of the author.  This software is distributed AS-IS.
15  */
16 
17 /*
18  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
19  */
20 
21 #include <sys/cdefs.h>
22 __FBSDID("$FreeBSD$");
23 
24 #include "opt_compat.h"
25 
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/protosw.h>
49 #include <sys/rwlock.h>
50 #include <sys/sema.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/syscall.h>
54 #include <sys/sysent.h>
55 #include <sys/sysctl.h>
56 #include <sys/syslog.h>
57 #include <sys/sx.h>
58 #include <sys/taskqueue.h>
59 #include <sys/vnode.h>
60 #include <sys/conf.h>
61 #include <sys/event.h>
62 #include <sys/mount.h>
63 #include <geom/geom.h>
64 
65 #include <machine/atomic.h>
66 
67 #include <vm/vm.h>
68 #include <vm/vm_page.h>
69 #include <vm/vm_extern.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
73 #include <vm/uma.h>
74 #include <sys/aio.h>
75 
76 /*
77  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
78  * overflow. (XXX will be removed soon.)
79  */
80 static u_long jobrefid;
81 
82 /*
83  * Counter for aio_fsync.
84  */
85 static uint64_t jobseqno;
86 
87 #ifndef MAX_AIO_PER_PROC
88 #define MAX_AIO_PER_PROC	32
89 #endif
90 
91 #ifndef MAX_AIO_QUEUE_PER_PROC
92 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
93 #endif
94 
95 #ifndef MAX_AIO_QUEUE
96 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
97 #endif
98 
99 #ifndef MAX_BUF_AIO
100 #define MAX_BUF_AIO		16
101 #endif
102 
103 FEATURE(aio, "Asynchronous I/O");
104 
105 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
106 
107 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
108     "Async IO management");
109 
110 static int enable_aio_unsafe = 0;
111 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
112     "Permit asynchronous IO on all file types, not just known-safe types");
113 
114 static unsigned int unsafe_warningcnt = 1;
115 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
116     &unsafe_warningcnt, 0,
117     "Warnings that will be triggered upon failed IO requests on unsafe files");
118 
119 static int max_aio_procs = MAX_AIO_PROCS;
120 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
121     "Maximum number of kernel processes to use for handling async IO ");
122 
123 static int num_aio_procs = 0;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
125     "Number of presently active kernel processes for async IO");
126 
127 /*
128  * The code will adjust the actual number of AIO processes towards this
129  * number when it gets a chance.
130  */
131 static int target_aio_procs = TARGET_AIO_PROCS;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
133     0,
134     "Preferred number of ready kernel processes for async IO");
135 
136 static int max_queue_count = MAX_AIO_QUEUE;
137 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
138     "Maximum number of aio requests to queue, globally");
139 
140 static int num_queue_count = 0;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
142     "Number of queued aio requests");
143 
144 static int num_buf_aio = 0;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
146     "Number of aio requests presently handled by the buf subsystem");
147 
148 /* Number of async I/O processes in the process of being started */
149 /* XXX This should be local to aio_aqueue() */
150 static int num_aio_resv_start = 0;
151 
152 static int aiod_lifetime;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
154     "Maximum lifetime for idle aiod");
155 
156 static int max_aio_per_proc = MAX_AIO_PER_PROC;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
158     0,
159     "Maximum active aio requests per process (stored in the process)");
160 
161 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
162 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
163     &max_aio_queue_per_proc, 0,
164     "Maximum queued aio requests per process (stored in the process)");
165 
166 static int max_buf_aio = MAX_BUF_AIO;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
168     "Maximum buf aio requests per process (stored in the process)");
169 
170 #ifdef COMPAT_FREEBSD6
171 typedef struct oaiocb {
172 	int	aio_fildes;		/* File descriptor */
173 	off_t	aio_offset;		/* File offset for I/O */
174 	volatile void *aio_buf;         /* I/O buffer in process space */
175 	size_t	aio_nbytes;		/* Number of bytes for I/O */
176 	struct	osigevent aio_sigevent;	/* Signal to deliver */
177 	int	aio_lio_opcode;		/* LIO opcode */
178 	int	aio_reqprio;		/* Request priority -- ignored */
179 	struct	__aiocb_private	_aiocb_private;
180 } oaiocb_t;
181 #endif
182 
183 /*
184  * Below is a key of locks used to protect each member of struct kaiocb
185  * aioliojob and kaioinfo and any backends.
186  *
187  * * - need not protected
188  * a - locked by kaioinfo lock
189  * b - locked by backend lock, the backend lock can be null in some cases,
190  *     for example, BIO belongs to this type, in this case, proc lock is
191  *     reused.
192  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
193  */
194 
195 /*
196  * If the routine that services an AIO request blocks while running in an
197  * AIO kernel process it can starve other I/O requests.  BIO requests
198  * queued via aio_qphysio() complete in GEOM and do not use AIO kernel
199  * processes at all.  Socket I/O requests use a separate pool of
200  * kprocs and also force non-blocking I/O.  Other file I/O requests
201  * use the generic fo_read/fo_write operations which can block.  The
202  * fsync and mlock operations can also block while executing.  Ideally
203  * none of these requests would block while executing.
204  *
205  * Note that the service routines cannot toggle O_NONBLOCK in the file
206  * structure directly while handling a request due to races with
207  * userland threads.
208  */
209 
210 /* jobflags */
211 #define	KAIOCB_QUEUEING		0x01
212 #define	KAIOCB_CANCELLED	0x02
213 #define	KAIOCB_CANCELLING	0x04
214 #define	KAIOCB_CHECKSYNC	0x08
215 #define	KAIOCB_CLEARED		0x10
216 #define	KAIOCB_FINISHED		0x20
217 
218 /*
219  * AIO process info
220  */
221 #define AIOP_FREE	0x1			/* proc on free queue */
222 
223 struct aioproc {
224 	int	aioprocflags;			/* (c) AIO proc flags */
225 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
226 	struct	proc *aioproc;			/* (*) the AIO proc */
227 };
228 
229 /*
230  * data-structure for lio signal management
231  */
232 struct aioliojob {
233 	int	lioj_flags;			/* (a) listio flags */
234 	int	lioj_count;			/* (a) listio flags */
235 	int	lioj_finished_count;		/* (a) listio flags */
236 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
237 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
238 	struct	knlist klist;			/* (a) list of knotes */
239 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
240 };
241 
242 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
243 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
244 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
245 
246 /*
247  * per process aio data structure
248  */
249 struct kaioinfo {
250 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
251 	int	kaio_flags;		/* (a) per process kaio flags */
252 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
253 	int	kaio_active_count;	/* (c) number of currently used AIOs */
254 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
255 	int	kaio_count;		/* (a) size of AIO queue */
256 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
257 	int	kaio_buffer_count;	/* (a) number of physio buffers */
258 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
259 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
260 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
261 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
262 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
263 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
264 	struct	task kaio_task;		/* (*) task to kick aio processes */
265 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
266 };
267 
268 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
269 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
270 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
271 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
272 
273 #define KAIO_RUNDOWN	0x1	/* process is being run down */
274 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
275 
276 /*
277  * Operations used to interact with userland aio control blocks.
278  * Different ABIs provide their own operations.
279  */
280 struct aiocb_ops {
281 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
282 	long	(*fetch_status)(struct aiocb *ujob);
283 	long	(*fetch_error)(struct aiocb *ujob);
284 	int	(*store_status)(struct aiocb *ujob, long status);
285 	int	(*store_error)(struct aiocb *ujob, long error);
286 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
287 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
288 };
289 
290 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
291 static struct sema aio_newproc_sem;
292 static struct mtx aio_job_mtx;
293 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
294 static struct unrhdr *aiod_unr;
295 
296 void		aio_init_aioinfo(struct proc *p);
297 static int	aio_onceonly(void);
298 static int	aio_free_entry(struct kaiocb *job);
299 static void	aio_process_rw(struct kaiocb *job);
300 static void	aio_process_sync(struct kaiocb *job);
301 static void	aio_process_mlock(struct kaiocb *job);
302 static void	aio_schedule_fsync(void *context, int pending);
303 static int	aio_newproc(int *);
304 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
305 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
306 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
307 static void	aio_physwakeup(struct bio *bp);
308 static void	aio_proc_rundown(void *arg, struct proc *p);
309 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
310 		    struct image_params *imgp);
311 static int	aio_qphysio(struct proc *p, struct kaiocb *job);
312 static void	aio_daemon(void *param);
313 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
314 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
315 static int	aio_kick(struct proc *userp);
316 static void	aio_kick_nowait(struct proc *userp);
317 static void	aio_kick_helper(void *context, int pending);
318 static int	filt_aioattach(struct knote *kn);
319 static void	filt_aiodetach(struct knote *kn);
320 static int	filt_aio(struct knote *kn, long hint);
321 static int	filt_lioattach(struct knote *kn);
322 static void	filt_liodetach(struct knote *kn);
323 static int	filt_lio(struct knote *kn, long hint);
324 
325 /*
326  * Zones for:
327  * 	kaio	Per process async io info
328  *	aiop	async io process data
329  *	aiocb	async io jobs
330  *	aiol	list io job pointer - internal to aio_suspend XXX
331  *	aiolio	list io jobs
332  */
333 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
334 
335 /* kqueue filters for aio */
336 static struct filterops aio_filtops = {
337 	.f_isfd = 0,
338 	.f_attach = filt_aioattach,
339 	.f_detach = filt_aiodetach,
340 	.f_event = filt_aio,
341 };
342 static struct filterops lio_filtops = {
343 	.f_isfd = 0,
344 	.f_attach = filt_lioattach,
345 	.f_detach = filt_liodetach,
346 	.f_event = filt_lio
347 };
348 
349 static eventhandler_tag exit_tag, exec_tag;
350 
351 TASKQUEUE_DEFINE_THREAD(aiod_kick);
352 
353 /*
354  * Main operations function for use as a kernel module.
355  */
356 static int
357 aio_modload(struct module *module, int cmd, void *arg)
358 {
359 	int error = 0;
360 
361 	switch (cmd) {
362 	case MOD_LOAD:
363 		aio_onceonly();
364 		break;
365 	case MOD_SHUTDOWN:
366 		break;
367 	default:
368 		error = EOPNOTSUPP;
369 		break;
370 	}
371 	return (error);
372 }
373 
374 static moduledata_t aio_mod = {
375 	"aio",
376 	&aio_modload,
377 	NULL
378 };
379 
380 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
381 MODULE_VERSION(aio, 1);
382 
383 /*
384  * Startup initialization
385  */
386 static int
387 aio_onceonly(void)
388 {
389 
390 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
391 	    EVENTHANDLER_PRI_ANY);
392 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
393 	    NULL, EVENTHANDLER_PRI_ANY);
394 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
395 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
396 	TAILQ_INIT(&aio_freeproc);
397 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
398 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
399 	TAILQ_INIT(&aio_jobs);
400 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
401 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
402 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
403 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
404 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
405 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
406 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
407 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
408 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
409 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
410 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
411 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
412 	jobrefid = 1;
413 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
414 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
415 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
416 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
417 
418 	return (0);
419 }
420 
421 /*
422  * Init the per-process aioinfo structure.  The aioinfo limits are set
423  * per-process for user limit (resource) management.
424  */
425 void
426 aio_init_aioinfo(struct proc *p)
427 {
428 	struct kaioinfo *ki;
429 
430 	ki = uma_zalloc(kaio_zone, M_WAITOK);
431 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
432 	ki->kaio_flags = 0;
433 	ki->kaio_maxactive_count = max_aio_per_proc;
434 	ki->kaio_active_count = 0;
435 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
436 	ki->kaio_count = 0;
437 	ki->kaio_ballowed_count = max_buf_aio;
438 	ki->kaio_buffer_count = 0;
439 	TAILQ_INIT(&ki->kaio_all);
440 	TAILQ_INIT(&ki->kaio_done);
441 	TAILQ_INIT(&ki->kaio_jobqueue);
442 	TAILQ_INIT(&ki->kaio_liojoblist);
443 	TAILQ_INIT(&ki->kaio_syncqueue);
444 	TAILQ_INIT(&ki->kaio_syncready);
445 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
446 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
447 	PROC_LOCK(p);
448 	if (p->p_aioinfo == NULL) {
449 		p->p_aioinfo = ki;
450 		PROC_UNLOCK(p);
451 	} else {
452 		PROC_UNLOCK(p);
453 		mtx_destroy(&ki->kaio_mtx);
454 		uma_zfree(kaio_zone, ki);
455 	}
456 
457 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
458 		aio_newproc(NULL);
459 }
460 
461 static int
462 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
463 {
464 	struct thread *td;
465 	int error;
466 
467 	error = sigev_findtd(p, sigev, &td);
468 	if (error)
469 		return (error);
470 	if (!KSI_ONQ(ksi)) {
471 		ksiginfo_set_sigev(ksi, sigev);
472 		ksi->ksi_code = SI_ASYNCIO;
473 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
474 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
475 	}
476 	PROC_UNLOCK(p);
477 	return (error);
478 }
479 
480 /*
481  * Free a job entry.  Wait for completion if it is currently active, but don't
482  * delay forever.  If we delay, we return a flag that says that we have to
483  * restart the queue scan.
484  */
485 static int
486 aio_free_entry(struct kaiocb *job)
487 {
488 	struct kaioinfo *ki;
489 	struct aioliojob *lj;
490 	struct proc *p;
491 
492 	p = job->userproc;
493 	MPASS(curproc == p);
494 	ki = p->p_aioinfo;
495 	MPASS(ki != NULL);
496 
497 	AIO_LOCK_ASSERT(ki, MA_OWNED);
498 	MPASS(job->jobflags & KAIOCB_FINISHED);
499 
500 	atomic_subtract_int(&num_queue_count, 1);
501 
502 	ki->kaio_count--;
503 	MPASS(ki->kaio_count >= 0);
504 
505 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
506 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
507 
508 	lj = job->lio;
509 	if (lj) {
510 		lj->lioj_count--;
511 		lj->lioj_finished_count--;
512 
513 		if (lj->lioj_count == 0) {
514 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
515 			/* lio is going away, we need to destroy any knotes */
516 			knlist_delete(&lj->klist, curthread, 1);
517 			PROC_LOCK(p);
518 			sigqueue_take(&lj->lioj_ksi);
519 			PROC_UNLOCK(p);
520 			uma_zfree(aiolio_zone, lj);
521 		}
522 	}
523 
524 	/* job is going away, we need to destroy any knotes */
525 	knlist_delete(&job->klist, curthread, 1);
526 	PROC_LOCK(p);
527 	sigqueue_take(&job->ksi);
528 	PROC_UNLOCK(p);
529 
530 	AIO_UNLOCK(ki);
531 
532 	/*
533 	 * The thread argument here is used to find the owning process
534 	 * and is also passed to fo_close() which may pass it to various
535 	 * places such as devsw close() routines.  Because of that, we
536 	 * need a thread pointer from the process owning the job that is
537 	 * persistent and won't disappear out from under us or move to
538 	 * another process.
539 	 *
540 	 * Currently, all the callers of this function call it to remove
541 	 * a kaiocb from the current process' job list either via a
542 	 * syscall or due to the current process calling exit() or
543 	 * execve().  Thus, we know that p == curproc.  We also know that
544 	 * curthread can't exit since we are curthread.
545 	 *
546 	 * Therefore, we use curthread as the thread to pass to
547 	 * knlist_delete().  This does mean that it is possible for the
548 	 * thread pointer at close time to differ from the thread pointer
549 	 * at open time, but this is already true of file descriptors in
550 	 * a multithreaded process.
551 	 */
552 	if (job->fd_file)
553 		fdrop(job->fd_file, curthread);
554 	crfree(job->cred);
555 	uma_zfree(aiocb_zone, job);
556 	AIO_LOCK(ki);
557 
558 	return (0);
559 }
560 
561 static void
562 aio_proc_rundown_exec(void *arg, struct proc *p,
563     struct image_params *imgp __unused)
564 {
565    	aio_proc_rundown(arg, p);
566 }
567 
568 static int
569 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
570 {
571 	aio_cancel_fn_t *func;
572 	int cancelled;
573 
574 	AIO_LOCK_ASSERT(ki, MA_OWNED);
575 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
576 		return (0);
577 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
578 	job->jobflags |= KAIOCB_CANCELLED;
579 
580 	func = job->cancel_fn;
581 
582 	/*
583 	 * If there is no cancel routine, just leave the job marked as
584 	 * cancelled.  The job should be in active use by a caller who
585 	 * should complete it normally or when it fails to install a
586 	 * cancel routine.
587 	 */
588 	if (func == NULL)
589 		return (0);
590 
591 	/*
592 	 * Set the CANCELLING flag so that aio_complete() will defer
593 	 * completions of this job.  This prevents the job from being
594 	 * freed out from under the cancel callback.  After the
595 	 * callback any deferred completion (whether from the callback
596 	 * or any other source) will be completed.
597 	 */
598 	job->jobflags |= KAIOCB_CANCELLING;
599 	AIO_UNLOCK(ki);
600 	func(job);
601 	AIO_LOCK(ki);
602 	job->jobflags &= ~KAIOCB_CANCELLING;
603 	if (job->jobflags & KAIOCB_FINISHED) {
604 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
605 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
606 		aio_bio_done_notify(p, job);
607 	} else {
608 		/*
609 		 * The cancel callback might have scheduled an
610 		 * operation to cancel this request, but it is
611 		 * only counted as cancelled if the request is
612 		 * cancelled when the callback returns.
613 		 */
614 		cancelled = 0;
615 	}
616 	return (cancelled);
617 }
618 
619 /*
620  * Rundown the jobs for a given process.
621  */
622 static void
623 aio_proc_rundown(void *arg, struct proc *p)
624 {
625 	struct kaioinfo *ki;
626 	struct aioliojob *lj;
627 	struct kaiocb *job, *jobn;
628 
629 	KASSERT(curthread->td_proc == p,
630 	    ("%s: called on non-curproc", __func__));
631 	ki = p->p_aioinfo;
632 	if (ki == NULL)
633 		return;
634 
635 	AIO_LOCK(ki);
636 	ki->kaio_flags |= KAIO_RUNDOWN;
637 
638 restart:
639 
640 	/*
641 	 * Try to cancel all pending requests. This code simulates
642 	 * aio_cancel on all pending I/O requests.
643 	 */
644 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
645 		aio_cancel_job(p, ki, job);
646 	}
647 
648 	/* Wait for all running I/O to be finished */
649 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
650 		ki->kaio_flags |= KAIO_WAKEUP;
651 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
652 		goto restart;
653 	}
654 
655 	/* Free all completed I/O requests. */
656 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
657 		aio_free_entry(job);
658 
659 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
660 		if (lj->lioj_count == 0) {
661 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
662 			knlist_delete(&lj->klist, curthread, 1);
663 			PROC_LOCK(p);
664 			sigqueue_take(&lj->lioj_ksi);
665 			PROC_UNLOCK(p);
666 			uma_zfree(aiolio_zone, lj);
667 		} else {
668 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
669 			    lj->lioj_count, lj->lioj_finished_count);
670 		}
671 	}
672 	AIO_UNLOCK(ki);
673 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
674 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
675 	mtx_destroy(&ki->kaio_mtx);
676 	uma_zfree(kaio_zone, ki);
677 	p->p_aioinfo = NULL;
678 }
679 
680 /*
681  * Select a job to run (called by an AIO daemon).
682  */
683 static struct kaiocb *
684 aio_selectjob(struct aioproc *aiop)
685 {
686 	struct kaiocb *job;
687 	struct kaioinfo *ki;
688 	struct proc *userp;
689 
690 	mtx_assert(&aio_job_mtx, MA_OWNED);
691 restart:
692 	TAILQ_FOREACH(job, &aio_jobs, list) {
693 		userp = job->userproc;
694 		ki = userp->p_aioinfo;
695 
696 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
697 			TAILQ_REMOVE(&aio_jobs, job, list);
698 			if (!aio_clear_cancel_function(job))
699 				goto restart;
700 
701 			/* Account for currently active jobs. */
702 			ki->kaio_active_count++;
703 			break;
704 		}
705 	}
706 	return (job);
707 }
708 
709 /*
710  * Move all data to a permanent storage device.  This code
711  * simulates the fsync syscall.
712  */
713 static int
714 aio_fsync_vnode(struct thread *td, struct vnode *vp)
715 {
716 	struct mount *mp;
717 	int error;
718 
719 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
720 		goto drop;
721 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
722 	if (vp->v_object != NULL) {
723 		VM_OBJECT_WLOCK(vp->v_object);
724 		vm_object_page_clean(vp->v_object, 0, 0, 0);
725 		VM_OBJECT_WUNLOCK(vp->v_object);
726 	}
727 	error = VOP_FSYNC(vp, MNT_WAIT, td);
728 
729 	VOP_UNLOCK(vp, 0);
730 	vn_finished_write(mp);
731 drop:
732 	return (error);
733 }
734 
735 /*
736  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
737  * does the I/O request for the non-physio version of the operations.  The
738  * normal vn operations are used, and this code should work in all instances
739  * for every type of file, including pipes, sockets, fifos, and regular files.
740  *
741  * XXX I don't think it works well for socket, pipe, and fifo.
742  */
743 static void
744 aio_process_rw(struct kaiocb *job)
745 {
746 	struct ucred *td_savedcred;
747 	struct thread *td;
748 	struct aiocb *cb;
749 	struct file *fp;
750 	struct uio auio;
751 	struct iovec aiov;
752 	ssize_t cnt;
753 	long msgsnd_st, msgsnd_end;
754 	long msgrcv_st, msgrcv_end;
755 	long oublock_st, oublock_end;
756 	long inblock_st, inblock_end;
757 	int error;
758 
759 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
760 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
761 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
762 
763 	aio_switch_vmspace(job);
764 	td = curthread;
765 	td_savedcred = td->td_ucred;
766 	td->td_ucred = job->cred;
767 	cb = &job->uaiocb;
768 	fp = job->fd_file;
769 
770 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
771 	aiov.iov_len = cb->aio_nbytes;
772 
773 	auio.uio_iov = &aiov;
774 	auio.uio_iovcnt = 1;
775 	auio.uio_offset = cb->aio_offset;
776 	auio.uio_resid = cb->aio_nbytes;
777 	cnt = cb->aio_nbytes;
778 	auio.uio_segflg = UIO_USERSPACE;
779 	auio.uio_td = td;
780 
781 	msgrcv_st = td->td_ru.ru_msgrcv;
782 	msgsnd_st = td->td_ru.ru_msgsnd;
783 	inblock_st = td->td_ru.ru_inblock;
784 	oublock_st = td->td_ru.ru_oublock;
785 
786 	/*
787 	 * aio_aqueue() acquires a reference to the file that is
788 	 * released in aio_free_entry().
789 	 */
790 	if (cb->aio_lio_opcode == LIO_READ) {
791 		auio.uio_rw = UIO_READ;
792 		if (auio.uio_resid == 0)
793 			error = 0;
794 		else
795 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
796 	} else {
797 		if (fp->f_type == DTYPE_VNODE)
798 			bwillwrite();
799 		auio.uio_rw = UIO_WRITE;
800 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
801 	}
802 	msgrcv_end = td->td_ru.ru_msgrcv;
803 	msgsnd_end = td->td_ru.ru_msgsnd;
804 	inblock_end = td->td_ru.ru_inblock;
805 	oublock_end = td->td_ru.ru_oublock;
806 
807 	job->msgrcv = msgrcv_end - msgrcv_st;
808 	job->msgsnd = msgsnd_end - msgsnd_st;
809 	job->inblock = inblock_end - inblock_st;
810 	job->outblock = oublock_end - oublock_st;
811 
812 	if ((error) && (auio.uio_resid != cnt)) {
813 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
814 			error = 0;
815 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
816 			PROC_LOCK(job->userproc);
817 			kern_psignal(job->userproc, SIGPIPE);
818 			PROC_UNLOCK(job->userproc);
819 		}
820 	}
821 
822 	cnt -= auio.uio_resid;
823 	td->td_ucred = td_savedcred;
824 	if (error)
825 		aio_complete(job, -1, error);
826 	else
827 		aio_complete(job, cnt, 0);
828 }
829 
830 static void
831 aio_process_sync(struct kaiocb *job)
832 {
833 	struct thread *td = curthread;
834 	struct ucred *td_savedcred = td->td_ucred;
835 	struct file *fp = job->fd_file;
836 	int error = 0;
837 
838 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
839 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
840 
841 	td->td_ucred = job->cred;
842 	if (fp->f_vnode != NULL)
843 		error = aio_fsync_vnode(td, fp->f_vnode);
844 	td->td_ucred = td_savedcred;
845 	if (error)
846 		aio_complete(job, -1, error);
847 	else
848 		aio_complete(job, 0, 0);
849 }
850 
851 static void
852 aio_process_mlock(struct kaiocb *job)
853 {
854 	struct aiocb *cb = &job->uaiocb;
855 	int error;
856 
857 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
858 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
859 
860 	aio_switch_vmspace(job);
861 	error = vm_mlock(job->userproc, job->cred,
862 	    __DEVOLATILE(void *, cb->aio_buf), cb->aio_nbytes);
863 	if (error)
864 		aio_complete(job, -1, error);
865 	else
866 		aio_complete(job, 0, 0);
867 }
868 
869 static void
870 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
871 {
872 	struct aioliojob *lj;
873 	struct kaioinfo *ki;
874 	struct kaiocb *sjob, *sjobn;
875 	int lj_done;
876 	bool schedule_fsync;
877 
878 	ki = userp->p_aioinfo;
879 	AIO_LOCK_ASSERT(ki, MA_OWNED);
880 	lj = job->lio;
881 	lj_done = 0;
882 	if (lj) {
883 		lj->lioj_finished_count++;
884 		if (lj->lioj_count == lj->lioj_finished_count)
885 			lj_done = 1;
886 	}
887 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
888 	MPASS(job->jobflags & KAIOCB_FINISHED);
889 
890 	if (ki->kaio_flags & KAIO_RUNDOWN)
891 		goto notification_done;
892 
893 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
894 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
895 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
896 
897 	KNOTE_LOCKED(&job->klist, 1);
898 
899 	if (lj_done) {
900 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
901 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
902 			KNOTE_LOCKED(&lj->klist, 1);
903 		}
904 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
905 		    == LIOJ_SIGNAL
906 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
907 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
908 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
909 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
910 		}
911 	}
912 
913 notification_done:
914 	if (job->jobflags & KAIOCB_CHECKSYNC) {
915 		schedule_fsync = false;
916 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
917 			if (job->fd_file != sjob->fd_file ||
918 			    job->seqno >= sjob->seqno)
919 				continue;
920 			if (--sjob->pending > 0)
921 				continue;
922 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
923 			if (!aio_clear_cancel_function_locked(sjob))
924 				continue;
925 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
926 			schedule_fsync = true;
927 		}
928 		if (schedule_fsync)
929 			taskqueue_enqueue(taskqueue_aiod_kick,
930 			    &ki->kaio_sync_task);
931 	}
932 	if (ki->kaio_flags & KAIO_WAKEUP) {
933 		ki->kaio_flags &= ~KAIO_WAKEUP;
934 		wakeup(&userp->p_aioinfo);
935 	}
936 }
937 
938 static void
939 aio_schedule_fsync(void *context, int pending)
940 {
941 	struct kaioinfo *ki;
942 	struct kaiocb *job;
943 
944 	ki = context;
945 	AIO_LOCK(ki);
946 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
947 		job = TAILQ_FIRST(&ki->kaio_syncready);
948 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
949 		AIO_UNLOCK(ki);
950 		aio_schedule(job, aio_process_sync);
951 		AIO_LOCK(ki);
952 	}
953 	AIO_UNLOCK(ki);
954 }
955 
956 bool
957 aio_cancel_cleared(struct kaiocb *job)
958 {
959 	struct kaioinfo *ki;
960 
961 	/*
962 	 * The caller should hold the same queue lock held when
963 	 * aio_clear_cancel_function() was called and set this flag
964 	 * ensuring this check sees an up-to-date value.  However,
965 	 * there is no way to assert that.
966 	 */
967 	ki = job->userproc->p_aioinfo;
968 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
969 }
970 
971 static bool
972 aio_clear_cancel_function_locked(struct kaiocb *job)
973 {
974 
975 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
976 	MPASS(job->cancel_fn != NULL);
977 	if (job->jobflags & KAIOCB_CANCELLING) {
978 		job->jobflags |= KAIOCB_CLEARED;
979 		return (false);
980 	}
981 	job->cancel_fn = NULL;
982 	return (true);
983 }
984 
985 bool
986 aio_clear_cancel_function(struct kaiocb *job)
987 {
988 	struct kaioinfo *ki;
989 	bool ret;
990 
991 	ki = job->userproc->p_aioinfo;
992 	AIO_LOCK(ki);
993 	ret = aio_clear_cancel_function_locked(job);
994 	AIO_UNLOCK(ki);
995 	return (ret);
996 }
997 
998 static bool
999 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1000 {
1001 
1002 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1003 	if (job->jobflags & KAIOCB_CANCELLED)
1004 		return (false);
1005 	job->cancel_fn = func;
1006 	return (true);
1007 }
1008 
1009 bool
1010 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1011 {
1012 	struct kaioinfo *ki;
1013 	bool ret;
1014 
1015 	ki = job->userproc->p_aioinfo;
1016 	AIO_LOCK(ki);
1017 	ret = aio_set_cancel_function_locked(job, func);
1018 	AIO_UNLOCK(ki);
1019 	return (ret);
1020 }
1021 
1022 void
1023 aio_complete(struct kaiocb *job, long status, int error)
1024 {
1025 	struct kaioinfo *ki;
1026 	struct proc *userp;
1027 
1028 	job->uaiocb._aiocb_private.error = error;
1029 	job->uaiocb._aiocb_private.status = status;
1030 
1031 	userp = job->userproc;
1032 	ki = userp->p_aioinfo;
1033 
1034 	AIO_LOCK(ki);
1035 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1036 	    ("duplicate aio_complete"));
1037 	job->jobflags |= KAIOCB_FINISHED;
1038 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1039 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1040 		aio_bio_done_notify(userp, job);
1041 	}
1042 	AIO_UNLOCK(ki);
1043 }
1044 
1045 void
1046 aio_cancel(struct kaiocb *job)
1047 {
1048 
1049 	aio_complete(job, -1, ECANCELED);
1050 }
1051 
1052 void
1053 aio_switch_vmspace(struct kaiocb *job)
1054 {
1055 
1056 	vmspace_switch_aio(job->userproc->p_vmspace);
1057 }
1058 
1059 /*
1060  * The AIO daemon, most of the actual work is done in aio_process_*,
1061  * but the setup (and address space mgmt) is done in this routine.
1062  */
1063 static void
1064 aio_daemon(void *_id)
1065 {
1066 	struct kaiocb *job;
1067 	struct aioproc *aiop;
1068 	struct kaioinfo *ki;
1069 	struct proc *p;
1070 	struct vmspace *myvm;
1071 	struct thread *td = curthread;
1072 	int id = (intptr_t)_id;
1073 
1074 	/*
1075 	 * Grab an extra reference on the daemon's vmspace so that it
1076 	 * doesn't get freed by jobs that switch to a different
1077 	 * vmspace.
1078 	 */
1079 	p = td->td_proc;
1080 	myvm = vmspace_acquire_ref(p);
1081 
1082 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1083 
1084 	/*
1085 	 * Allocate and ready the aio control info.  There is one aiop structure
1086 	 * per daemon.
1087 	 */
1088 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1089 	aiop->aioproc = p;
1090 	aiop->aioprocflags = 0;
1091 
1092 	/*
1093 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1094 	 * and creating too many daemons.)
1095 	 */
1096 	sema_post(&aio_newproc_sem);
1097 
1098 	mtx_lock(&aio_job_mtx);
1099 	for (;;) {
1100 		/*
1101 		 * Take daemon off of free queue
1102 		 */
1103 		if (aiop->aioprocflags & AIOP_FREE) {
1104 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1105 			aiop->aioprocflags &= ~AIOP_FREE;
1106 		}
1107 
1108 		/*
1109 		 * Check for jobs.
1110 		 */
1111 		while ((job = aio_selectjob(aiop)) != NULL) {
1112 			mtx_unlock(&aio_job_mtx);
1113 
1114 			ki = job->userproc->p_aioinfo;
1115 			job->handle_fn(job);
1116 
1117 			mtx_lock(&aio_job_mtx);
1118 			/* Decrement the active job count. */
1119 			ki->kaio_active_count--;
1120 		}
1121 
1122 		/*
1123 		 * Disconnect from user address space.
1124 		 */
1125 		if (p->p_vmspace != myvm) {
1126 			mtx_unlock(&aio_job_mtx);
1127 			vmspace_switch_aio(myvm);
1128 			mtx_lock(&aio_job_mtx);
1129 			/*
1130 			 * We have to restart to avoid race, we only sleep if
1131 			 * no job can be selected.
1132 			 */
1133 			continue;
1134 		}
1135 
1136 		mtx_assert(&aio_job_mtx, MA_OWNED);
1137 
1138 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1139 		aiop->aioprocflags |= AIOP_FREE;
1140 
1141 		/*
1142 		 * If daemon is inactive for a long time, allow it to exit,
1143 		 * thereby freeing resources.
1144 		 */
1145 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1146 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1147 		    (aiop->aioprocflags & AIOP_FREE) &&
1148 		    num_aio_procs > target_aio_procs)
1149 			break;
1150 	}
1151 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1152 	num_aio_procs--;
1153 	mtx_unlock(&aio_job_mtx);
1154 	uma_zfree(aiop_zone, aiop);
1155 	free_unr(aiod_unr, id);
1156 	vmspace_free(myvm);
1157 
1158 	KASSERT(p->p_vmspace == myvm,
1159 	    ("AIOD: bad vmspace for exiting daemon"));
1160 	KASSERT(myvm->vm_refcnt > 1,
1161 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1162 	kproc_exit(0);
1163 }
1164 
1165 /*
1166  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1167  * AIO daemon modifies its environment itself.
1168  */
1169 static int
1170 aio_newproc(int *start)
1171 {
1172 	int error;
1173 	struct proc *p;
1174 	int id;
1175 
1176 	id = alloc_unr(aiod_unr);
1177 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1178 		RFNOWAIT, 0, "aiod%d", id);
1179 	if (error == 0) {
1180 		/*
1181 		 * Wait until daemon is started.
1182 		 */
1183 		sema_wait(&aio_newproc_sem);
1184 		mtx_lock(&aio_job_mtx);
1185 		num_aio_procs++;
1186 		if (start != NULL)
1187 			(*start)--;
1188 		mtx_unlock(&aio_job_mtx);
1189 	} else {
1190 		free_unr(aiod_unr, id);
1191 	}
1192 	return (error);
1193 }
1194 
1195 /*
1196  * Try the high-performance, low-overhead physio method for eligible
1197  * VCHR devices.  This method doesn't use an aio helper thread, and
1198  * thus has very low overhead.
1199  *
1200  * Assumes that the caller, aio_aqueue(), has incremented the file
1201  * structure's reference count, preventing its deallocation for the
1202  * duration of this call.
1203  */
1204 static int
1205 aio_qphysio(struct proc *p, struct kaiocb *job)
1206 {
1207 	struct aiocb *cb;
1208 	struct file *fp;
1209 	struct bio *bp;
1210 	struct buf *pbuf;
1211 	struct vnode *vp;
1212 	struct cdevsw *csw;
1213 	struct cdev *dev;
1214 	struct kaioinfo *ki;
1215 	int error, ref, poff;
1216 	vm_prot_t prot;
1217 
1218 	cb = &job->uaiocb;
1219 	fp = job->fd_file;
1220 
1221 	if (fp == NULL || fp->f_type != DTYPE_VNODE)
1222 		return (-1);
1223 
1224 	vp = fp->f_vnode;
1225 	if (vp->v_type != VCHR)
1226 		return (-1);
1227 	if (vp->v_bufobj.bo_bsize == 0)
1228 		return (-1);
1229 	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1230 		return (-1);
1231 
1232 	ref = 0;
1233 	csw = devvn_refthread(vp, &dev, &ref);
1234 	if (csw == NULL)
1235 		return (ENXIO);
1236 
1237 	if ((csw->d_flags & D_DISK) == 0) {
1238 		error = -1;
1239 		goto unref;
1240 	}
1241 	if (cb->aio_nbytes > dev->si_iosize_max) {
1242 		error = -1;
1243 		goto unref;
1244 	}
1245 
1246 	ki = p->p_aioinfo;
1247 	poff = (vm_offset_t)cb->aio_buf & PAGE_MASK;
1248 	if ((dev->si_flags & SI_UNMAPPED) && unmapped_buf_allowed) {
1249 		if (cb->aio_nbytes > MAXPHYS) {
1250 			error = -1;
1251 			goto unref;
1252 		}
1253 
1254 		pbuf = NULL;
1255 	} else {
1256 		if (cb->aio_nbytes > MAXPHYS - poff) {
1257 			error = -1;
1258 			goto unref;
1259 		}
1260 		if (ki->kaio_buffer_count >= ki->kaio_ballowed_count) {
1261 			error = -1;
1262 			goto unref;
1263 		}
1264 
1265 		job->pbuf = pbuf = (struct buf *)getpbuf(NULL);
1266 		BUF_KERNPROC(pbuf);
1267 		AIO_LOCK(ki);
1268 		ki->kaio_buffer_count++;
1269 		AIO_UNLOCK(ki);
1270 	}
1271 	job->bp = bp = g_alloc_bio();
1272 
1273 	bp->bio_length = cb->aio_nbytes;
1274 	bp->bio_bcount = cb->aio_nbytes;
1275 	bp->bio_done = aio_physwakeup;
1276 	bp->bio_data = (void *)(uintptr_t)cb->aio_buf;
1277 	bp->bio_offset = cb->aio_offset;
1278 	bp->bio_cmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1279 	bp->bio_dev = dev;
1280 	bp->bio_caller1 = (void *)job;
1281 
1282 	prot = VM_PROT_READ;
1283 	if (cb->aio_lio_opcode == LIO_READ)
1284 		prot |= VM_PROT_WRITE;	/* Less backwards than it looks */
1285 	job->npages = vm_fault_quick_hold_pages(&curproc->p_vmspace->vm_map,
1286 	    (vm_offset_t)bp->bio_data, bp->bio_length, prot, job->pages,
1287 	    nitems(job->pages));
1288 	if (job->npages < 0) {
1289 		error = EFAULT;
1290 		goto doerror;
1291 	}
1292 	if (pbuf != NULL) {
1293 		pmap_qenter((vm_offset_t)pbuf->b_data,
1294 		    job->pages, job->npages);
1295 		bp->bio_data = pbuf->b_data + poff;
1296 		atomic_add_int(&num_buf_aio, 1);
1297 	} else {
1298 		bp->bio_ma = job->pages;
1299 		bp->bio_ma_n = job->npages;
1300 		bp->bio_ma_offset = poff;
1301 		bp->bio_data = unmapped_buf;
1302 		bp->bio_flags |= BIO_UNMAPPED;
1303 	}
1304 
1305 	/* Perform transfer. */
1306 	csw->d_strategy(bp);
1307 	dev_relthread(dev, ref);
1308 	return (0);
1309 
1310 doerror:
1311 	if (pbuf != NULL) {
1312 		AIO_LOCK(ki);
1313 		ki->kaio_buffer_count--;
1314 		AIO_UNLOCK(ki);
1315 		relpbuf(pbuf, NULL);
1316 		job->pbuf = NULL;
1317 	}
1318 	g_destroy_bio(bp);
1319 	job->bp = NULL;
1320 unref:
1321 	dev_relthread(dev, ref);
1322 	return (error);
1323 }
1324 
1325 #ifdef COMPAT_FREEBSD6
1326 static int
1327 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1328 {
1329 
1330 	/*
1331 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1332 	 * supported by AIO with the old sigevent structure.
1333 	 */
1334 	nsig->sigev_notify = osig->sigev_notify;
1335 	switch (nsig->sigev_notify) {
1336 	case SIGEV_NONE:
1337 		break;
1338 	case SIGEV_SIGNAL:
1339 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1340 		break;
1341 	case SIGEV_KEVENT:
1342 		nsig->sigev_notify_kqueue =
1343 		    osig->__sigev_u.__sigev_notify_kqueue;
1344 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1345 		break;
1346 	default:
1347 		return (EINVAL);
1348 	}
1349 	return (0);
1350 }
1351 
1352 static int
1353 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1354 {
1355 	struct oaiocb *ojob;
1356 	int error;
1357 
1358 	bzero(kjob, sizeof(struct aiocb));
1359 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1360 	if (error)
1361 		return (error);
1362 	ojob = (struct oaiocb *)kjob;
1363 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1364 }
1365 #endif
1366 
1367 static int
1368 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1369 {
1370 
1371 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1372 }
1373 
1374 static long
1375 aiocb_fetch_status(struct aiocb *ujob)
1376 {
1377 
1378 	return (fuword(&ujob->_aiocb_private.status));
1379 }
1380 
1381 static long
1382 aiocb_fetch_error(struct aiocb *ujob)
1383 {
1384 
1385 	return (fuword(&ujob->_aiocb_private.error));
1386 }
1387 
1388 static int
1389 aiocb_store_status(struct aiocb *ujob, long status)
1390 {
1391 
1392 	return (suword(&ujob->_aiocb_private.status, status));
1393 }
1394 
1395 static int
1396 aiocb_store_error(struct aiocb *ujob, long error)
1397 {
1398 
1399 	return (suword(&ujob->_aiocb_private.error, error));
1400 }
1401 
1402 static int
1403 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1404 {
1405 
1406 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1407 }
1408 
1409 static int
1410 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1411 {
1412 
1413 	return (suword(ujobp, (long)ujob));
1414 }
1415 
1416 static struct aiocb_ops aiocb_ops = {
1417 	.copyin = aiocb_copyin,
1418 	.fetch_status = aiocb_fetch_status,
1419 	.fetch_error = aiocb_fetch_error,
1420 	.store_status = aiocb_store_status,
1421 	.store_error = aiocb_store_error,
1422 	.store_kernelinfo = aiocb_store_kernelinfo,
1423 	.store_aiocb = aiocb_store_aiocb,
1424 };
1425 
1426 #ifdef COMPAT_FREEBSD6
1427 static struct aiocb_ops aiocb_ops_osigevent = {
1428 	.copyin = aiocb_copyin_old_sigevent,
1429 	.fetch_status = aiocb_fetch_status,
1430 	.fetch_error = aiocb_fetch_error,
1431 	.store_status = aiocb_store_status,
1432 	.store_error = aiocb_store_error,
1433 	.store_kernelinfo = aiocb_store_kernelinfo,
1434 	.store_aiocb = aiocb_store_aiocb,
1435 };
1436 #endif
1437 
1438 /*
1439  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1440  * technique is done in this code.
1441  */
1442 int
1443 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1444     int type, struct aiocb_ops *ops)
1445 {
1446 	struct proc *p = td->td_proc;
1447 	cap_rights_t rights;
1448 	struct file *fp;
1449 	struct kaiocb *job;
1450 	struct kaioinfo *ki;
1451 	struct kevent kev;
1452 	int opcode;
1453 	int error;
1454 	int fd, kqfd;
1455 	int jid;
1456 	u_short evflags;
1457 
1458 	if (p->p_aioinfo == NULL)
1459 		aio_init_aioinfo(p);
1460 
1461 	ki = p->p_aioinfo;
1462 
1463 	ops->store_status(ujob, -1);
1464 	ops->store_error(ujob, 0);
1465 	ops->store_kernelinfo(ujob, -1);
1466 
1467 	if (num_queue_count >= max_queue_count ||
1468 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1469 		ops->store_error(ujob, EAGAIN);
1470 		return (EAGAIN);
1471 	}
1472 
1473 	job = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1474 	knlist_init_mtx(&job->klist, AIO_MTX(ki));
1475 
1476 	error = ops->copyin(ujob, &job->uaiocb);
1477 	if (error) {
1478 		ops->store_error(ujob, error);
1479 		uma_zfree(aiocb_zone, job);
1480 		return (error);
1481 	}
1482 
1483 	if (job->uaiocb.aio_nbytes > IOSIZE_MAX) {
1484 		uma_zfree(aiocb_zone, job);
1485 		return (EINVAL);
1486 	}
1487 
1488 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1489 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1490 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1491 	    job->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1492 		ops->store_error(ujob, EINVAL);
1493 		uma_zfree(aiocb_zone, job);
1494 		return (EINVAL);
1495 	}
1496 
1497 	if ((job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1498 	     job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1499 		!_SIG_VALID(job->uaiocb.aio_sigevent.sigev_signo)) {
1500 		uma_zfree(aiocb_zone, job);
1501 		return (EINVAL);
1502 	}
1503 
1504 	ksiginfo_init(&job->ksi);
1505 
1506 	/* Save userspace address of the job info. */
1507 	job->ujob = ujob;
1508 
1509 	/* Get the opcode. */
1510 	if (type != LIO_NOP)
1511 		job->uaiocb.aio_lio_opcode = type;
1512 	opcode = job->uaiocb.aio_lio_opcode;
1513 
1514 	/*
1515 	 * Validate the opcode and fetch the file object for the specified
1516 	 * file descriptor.
1517 	 *
1518 	 * XXXRW: Moved the opcode validation up here so that we don't
1519 	 * retrieve a file descriptor without knowing what the capabiltity
1520 	 * should be.
1521 	 */
1522 	fd = job->uaiocb.aio_fildes;
1523 	switch (opcode) {
1524 	case LIO_WRITE:
1525 		error = fget_write(td, fd,
1526 		    cap_rights_init(&rights, CAP_PWRITE), &fp);
1527 		break;
1528 	case LIO_READ:
1529 		error = fget_read(td, fd,
1530 		    cap_rights_init(&rights, CAP_PREAD), &fp);
1531 		break;
1532 	case LIO_SYNC:
1533 		error = fget(td, fd, cap_rights_init(&rights, CAP_FSYNC), &fp);
1534 		break;
1535 	case LIO_MLOCK:
1536 		fp = NULL;
1537 		break;
1538 	case LIO_NOP:
1539 		error = fget(td, fd, cap_rights_init(&rights), &fp);
1540 		break;
1541 	default:
1542 		error = EINVAL;
1543 	}
1544 	if (error) {
1545 		uma_zfree(aiocb_zone, job);
1546 		ops->store_error(ujob, error);
1547 		return (error);
1548 	}
1549 
1550 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1551 		error = EINVAL;
1552 		goto aqueue_fail;
1553 	}
1554 
1555 	if (opcode != LIO_SYNC && job->uaiocb.aio_offset == -1LL) {
1556 		error = EINVAL;
1557 		goto aqueue_fail;
1558 	}
1559 
1560 	job->fd_file = fp;
1561 
1562 	mtx_lock(&aio_job_mtx);
1563 	jid = jobrefid++;
1564 	job->seqno = jobseqno++;
1565 	mtx_unlock(&aio_job_mtx);
1566 	error = ops->store_kernelinfo(ujob, jid);
1567 	if (error) {
1568 		error = EINVAL;
1569 		goto aqueue_fail;
1570 	}
1571 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1572 
1573 	if (opcode == LIO_NOP) {
1574 		fdrop(fp, td);
1575 		uma_zfree(aiocb_zone, job);
1576 		return (0);
1577 	}
1578 
1579 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1580 		goto no_kqueue;
1581 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1582 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1583 		error = EINVAL;
1584 		goto aqueue_fail;
1585 	}
1586 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1587 	kev.ident = (uintptr_t)job->ujob;
1588 	kev.filter = EVFILT_AIO;
1589 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1590 	kev.data = (intptr_t)job;
1591 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1592 	error = kqfd_register(kqfd, &kev, td, 1);
1593 	if (error)
1594 		goto aqueue_fail;
1595 
1596 no_kqueue:
1597 
1598 	ops->store_error(ujob, EINPROGRESS);
1599 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1600 	job->userproc = p;
1601 	job->cred = crhold(td->td_ucred);
1602 	job->jobflags = KAIOCB_QUEUEING;
1603 	job->lio = lj;
1604 
1605 	if (opcode == LIO_MLOCK) {
1606 		aio_schedule(job, aio_process_mlock);
1607 		error = 0;
1608 	} else if (fp->f_ops->fo_aio_queue == NULL)
1609 		error = aio_queue_file(fp, job);
1610 	else
1611 		error = fo_aio_queue(fp, job);
1612 	if (error)
1613 		goto aqueue_fail;
1614 
1615 	AIO_LOCK(ki);
1616 	job->jobflags &= ~KAIOCB_QUEUEING;
1617 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1618 	ki->kaio_count++;
1619 	if (lj)
1620 		lj->lioj_count++;
1621 	atomic_add_int(&num_queue_count, 1);
1622 	if (job->jobflags & KAIOCB_FINISHED) {
1623 		/*
1624 		 * The queue callback completed the request synchronously.
1625 		 * The bulk of the completion is deferred in that case
1626 		 * until this point.
1627 		 */
1628 		aio_bio_done_notify(p, job);
1629 	} else
1630 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1631 	AIO_UNLOCK(ki);
1632 	return (0);
1633 
1634 aqueue_fail:
1635 	knlist_delete(&job->klist, curthread, 0);
1636 	if (fp)
1637 		fdrop(fp, td);
1638 	uma_zfree(aiocb_zone, job);
1639 	ops->store_error(ujob, error);
1640 	return (error);
1641 }
1642 
1643 static void
1644 aio_cancel_daemon_job(struct kaiocb *job)
1645 {
1646 
1647 	mtx_lock(&aio_job_mtx);
1648 	if (!aio_cancel_cleared(job))
1649 		TAILQ_REMOVE(&aio_jobs, job, list);
1650 	mtx_unlock(&aio_job_mtx);
1651 	aio_cancel(job);
1652 }
1653 
1654 void
1655 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1656 {
1657 
1658 	mtx_lock(&aio_job_mtx);
1659 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1660 		mtx_unlock(&aio_job_mtx);
1661 		aio_cancel(job);
1662 		return;
1663 	}
1664 	job->handle_fn = func;
1665 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1666 	aio_kick_nowait(job->userproc);
1667 	mtx_unlock(&aio_job_mtx);
1668 }
1669 
1670 static void
1671 aio_cancel_sync(struct kaiocb *job)
1672 {
1673 	struct kaioinfo *ki;
1674 
1675 	ki = job->userproc->p_aioinfo;
1676 	AIO_LOCK(ki);
1677 	if (!aio_cancel_cleared(job))
1678 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1679 	AIO_UNLOCK(ki);
1680 	aio_cancel(job);
1681 }
1682 
1683 int
1684 aio_queue_file(struct file *fp, struct kaiocb *job)
1685 {
1686 	struct aioliojob *lj;
1687 	struct kaioinfo *ki;
1688 	struct kaiocb *job2;
1689 	struct vnode *vp;
1690 	struct mount *mp;
1691 	int error, opcode;
1692 	bool safe;
1693 
1694 	lj = job->lio;
1695 	ki = job->userproc->p_aioinfo;
1696 	opcode = job->uaiocb.aio_lio_opcode;
1697 	if (opcode == LIO_SYNC)
1698 		goto queueit;
1699 
1700 	if ((error = aio_qphysio(job->userproc, job)) == 0)
1701 		goto done;
1702 #if 0
1703 	/*
1704 	 * XXX: This means qphysio() failed with EFAULT.  The current
1705 	 * behavior is to retry the operation via fo_read/fo_write.
1706 	 * Wouldn't it be better to just complete the request with an
1707 	 * error here?
1708 	 */
1709 	if (error > 0)
1710 		goto done;
1711 #endif
1712 queueit:
1713 	safe = false;
1714 	if (fp->f_type == DTYPE_VNODE) {
1715 		vp = fp->f_vnode;
1716 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1717 			mp = fp->f_vnode->v_mount;
1718 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1719 				safe = true;
1720 		}
1721 	}
1722 	if (!(safe || enable_aio_unsafe)) {
1723 		counted_warning(&unsafe_warningcnt,
1724 		    "is attempting to use unsafe AIO requests");
1725 		return (EOPNOTSUPP);
1726 	}
1727 
1728 	if (opcode == LIO_SYNC) {
1729 		AIO_LOCK(ki);
1730 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1731 			if (job2->fd_file == job->fd_file &&
1732 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1733 			    job2->seqno < job->seqno) {
1734 				job2->jobflags |= KAIOCB_CHECKSYNC;
1735 				job->pending++;
1736 			}
1737 		}
1738 		if (job->pending != 0) {
1739 			if (!aio_set_cancel_function_locked(job,
1740 				aio_cancel_sync)) {
1741 				AIO_UNLOCK(ki);
1742 				aio_cancel(job);
1743 				return (0);
1744 			}
1745 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1746 			AIO_UNLOCK(ki);
1747 			return (0);
1748 		}
1749 		AIO_UNLOCK(ki);
1750 	}
1751 
1752 	switch (opcode) {
1753 	case LIO_READ:
1754 	case LIO_WRITE:
1755 		aio_schedule(job, aio_process_rw);
1756 		error = 0;
1757 		break;
1758 	case LIO_SYNC:
1759 		aio_schedule(job, aio_process_sync);
1760 		error = 0;
1761 		break;
1762 	default:
1763 		error = EINVAL;
1764 	}
1765 done:
1766 	return (error);
1767 }
1768 
1769 static void
1770 aio_kick_nowait(struct proc *userp)
1771 {
1772 	struct kaioinfo *ki = userp->p_aioinfo;
1773 	struct aioproc *aiop;
1774 
1775 	mtx_assert(&aio_job_mtx, MA_OWNED);
1776 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1777 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1778 		aiop->aioprocflags &= ~AIOP_FREE;
1779 		wakeup(aiop->aioproc);
1780 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1781 	    ki->kaio_active_count + num_aio_resv_start <
1782 	    ki->kaio_maxactive_count) {
1783 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1784 	}
1785 }
1786 
1787 static int
1788 aio_kick(struct proc *userp)
1789 {
1790 	struct kaioinfo *ki = userp->p_aioinfo;
1791 	struct aioproc *aiop;
1792 	int error, ret = 0;
1793 
1794 	mtx_assert(&aio_job_mtx, MA_OWNED);
1795 retryproc:
1796 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1797 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1798 		aiop->aioprocflags &= ~AIOP_FREE;
1799 		wakeup(aiop->aioproc);
1800 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1801 	    ki->kaio_active_count + num_aio_resv_start <
1802 	    ki->kaio_maxactive_count) {
1803 		num_aio_resv_start++;
1804 		mtx_unlock(&aio_job_mtx);
1805 		error = aio_newproc(&num_aio_resv_start);
1806 		mtx_lock(&aio_job_mtx);
1807 		if (error) {
1808 			num_aio_resv_start--;
1809 			goto retryproc;
1810 		}
1811 	} else {
1812 		ret = -1;
1813 	}
1814 	return (ret);
1815 }
1816 
1817 static void
1818 aio_kick_helper(void *context, int pending)
1819 {
1820 	struct proc *userp = context;
1821 
1822 	mtx_lock(&aio_job_mtx);
1823 	while (--pending >= 0) {
1824 		if (aio_kick(userp))
1825 			break;
1826 	}
1827 	mtx_unlock(&aio_job_mtx);
1828 }
1829 
1830 /*
1831  * Support the aio_return system call, as a side-effect, kernel resources are
1832  * released.
1833  */
1834 static int
1835 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1836 {
1837 	struct proc *p = td->td_proc;
1838 	struct kaiocb *job;
1839 	struct kaioinfo *ki;
1840 	long status, error;
1841 
1842 	ki = p->p_aioinfo;
1843 	if (ki == NULL)
1844 		return (EINVAL);
1845 	AIO_LOCK(ki);
1846 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1847 		if (job->ujob == ujob)
1848 			break;
1849 	}
1850 	if (job != NULL) {
1851 		MPASS(job->jobflags & KAIOCB_FINISHED);
1852 		status = job->uaiocb._aiocb_private.status;
1853 		error = job->uaiocb._aiocb_private.error;
1854 		td->td_retval[0] = status;
1855 		td->td_ru.ru_oublock += job->outblock;
1856 		td->td_ru.ru_inblock += job->inblock;
1857 		td->td_ru.ru_msgsnd += job->msgsnd;
1858 		td->td_ru.ru_msgrcv += job->msgrcv;
1859 		aio_free_entry(job);
1860 		AIO_UNLOCK(ki);
1861 		ops->store_error(ujob, error);
1862 		ops->store_status(ujob, status);
1863 	} else {
1864 		error = EINVAL;
1865 		AIO_UNLOCK(ki);
1866 	}
1867 	return (error);
1868 }
1869 
1870 int
1871 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1872 {
1873 
1874 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1875 }
1876 
1877 /*
1878  * Allow a process to wakeup when any of the I/O requests are completed.
1879  */
1880 static int
1881 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1882     struct timespec *ts)
1883 {
1884 	struct proc *p = td->td_proc;
1885 	struct timeval atv;
1886 	struct kaioinfo *ki;
1887 	struct kaiocb *firstjob, *job;
1888 	int error, i, timo;
1889 
1890 	timo = 0;
1891 	if (ts) {
1892 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1893 			return (EINVAL);
1894 
1895 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1896 		if (itimerfix(&atv))
1897 			return (EINVAL);
1898 		timo = tvtohz(&atv);
1899 	}
1900 
1901 	ki = p->p_aioinfo;
1902 	if (ki == NULL)
1903 		return (EAGAIN);
1904 
1905 	if (njoblist == 0)
1906 		return (0);
1907 
1908 	AIO_LOCK(ki);
1909 	for (;;) {
1910 		firstjob = NULL;
1911 		error = 0;
1912 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1913 			for (i = 0; i < njoblist; i++) {
1914 				if (job->ujob == ujoblist[i]) {
1915 					if (firstjob == NULL)
1916 						firstjob = job;
1917 					if (job->jobflags & KAIOCB_FINISHED)
1918 						goto RETURN;
1919 				}
1920 			}
1921 		}
1922 		/* All tasks were finished. */
1923 		if (firstjob == NULL)
1924 			break;
1925 
1926 		ki->kaio_flags |= KAIO_WAKEUP;
1927 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1928 		    "aiospn", timo);
1929 		if (error == ERESTART)
1930 			error = EINTR;
1931 		if (error)
1932 			break;
1933 	}
1934 RETURN:
1935 	AIO_UNLOCK(ki);
1936 	return (error);
1937 }
1938 
1939 int
1940 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1941 {
1942 	struct timespec ts, *tsp;
1943 	struct aiocb **ujoblist;
1944 	int error;
1945 
1946 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1947 		return (EINVAL);
1948 
1949 	if (uap->timeout) {
1950 		/* Get timespec struct. */
1951 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1952 			return (error);
1953 		tsp = &ts;
1954 	} else
1955 		tsp = NULL;
1956 
1957 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1958 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1959 	if (error == 0)
1960 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1961 	uma_zfree(aiol_zone, ujoblist);
1962 	return (error);
1963 }
1964 
1965 /*
1966  * aio_cancel cancels any non-physio aio operations not currently in
1967  * progress.
1968  */
1969 int
1970 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1971 {
1972 	struct proc *p = td->td_proc;
1973 	struct kaioinfo *ki;
1974 	struct kaiocb *job, *jobn;
1975 	struct file *fp;
1976 	cap_rights_t rights;
1977 	int error;
1978 	int cancelled = 0;
1979 	int notcancelled = 0;
1980 	struct vnode *vp;
1981 
1982 	/* Lookup file object. */
1983 	error = fget(td, uap->fd, cap_rights_init(&rights), &fp);
1984 	if (error)
1985 		return (error);
1986 
1987 	ki = p->p_aioinfo;
1988 	if (ki == NULL)
1989 		goto done;
1990 
1991 	if (fp->f_type == DTYPE_VNODE) {
1992 		vp = fp->f_vnode;
1993 		if (vn_isdisk(vp, &error)) {
1994 			fdrop(fp, td);
1995 			td->td_retval[0] = AIO_NOTCANCELED;
1996 			return (0);
1997 		}
1998 	}
1999 
2000 	AIO_LOCK(ki);
2001 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
2002 		if ((uap->fd == job->uaiocb.aio_fildes) &&
2003 		    ((uap->aiocbp == NULL) ||
2004 		     (uap->aiocbp == job->ujob))) {
2005 			if (aio_cancel_job(p, ki, job)) {
2006 				cancelled++;
2007 			} else {
2008 				notcancelled++;
2009 			}
2010 			if (uap->aiocbp != NULL)
2011 				break;
2012 		}
2013 	}
2014 	AIO_UNLOCK(ki);
2015 
2016 done:
2017 	fdrop(fp, td);
2018 
2019 	if (uap->aiocbp != NULL) {
2020 		if (cancelled) {
2021 			td->td_retval[0] = AIO_CANCELED;
2022 			return (0);
2023 		}
2024 	}
2025 
2026 	if (notcancelled) {
2027 		td->td_retval[0] = AIO_NOTCANCELED;
2028 		return (0);
2029 	}
2030 
2031 	if (cancelled) {
2032 		td->td_retval[0] = AIO_CANCELED;
2033 		return (0);
2034 	}
2035 
2036 	td->td_retval[0] = AIO_ALLDONE;
2037 
2038 	return (0);
2039 }
2040 
2041 /*
2042  * aio_error is implemented in the kernel level for compatibility purposes
2043  * only.  For a user mode async implementation, it would be best to do it in
2044  * a userland subroutine.
2045  */
2046 static int
2047 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2048 {
2049 	struct proc *p = td->td_proc;
2050 	struct kaiocb *job;
2051 	struct kaioinfo *ki;
2052 	int status;
2053 
2054 	ki = p->p_aioinfo;
2055 	if (ki == NULL) {
2056 		td->td_retval[0] = EINVAL;
2057 		return (0);
2058 	}
2059 
2060 	AIO_LOCK(ki);
2061 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2062 		if (job->ujob == ujob) {
2063 			if (job->jobflags & KAIOCB_FINISHED)
2064 				td->td_retval[0] =
2065 					job->uaiocb._aiocb_private.error;
2066 			else
2067 				td->td_retval[0] = EINPROGRESS;
2068 			AIO_UNLOCK(ki);
2069 			return (0);
2070 		}
2071 	}
2072 	AIO_UNLOCK(ki);
2073 
2074 	/*
2075 	 * Hack for failure of aio_aqueue.
2076 	 */
2077 	status = ops->fetch_status(ujob);
2078 	if (status == -1) {
2079 		td->td_retval[0] = ops->fetch_error(ujob);
2080 		return (0);
2081 	}
2082 
2083 	td->td_retval[0] = EINVAL;
2084 	return (0);
2085 }
2086 
2087 int
2088 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2089 {
2090 
2091 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2092 }
2093 
2094 /* syscall - asynchronous read from a file (REALTIME) */
2095 #ifdef COMPAT_FREEBSD6
2096 int
2097 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2098 {
2099 
2100 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2101 	    &aiocb_ops_osigevent));
2102 }
2103 #endif
2104 
2105 int
2106 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2107 {
2108 
2109 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2110 }
2111 
2112 /* syscall - asynchronous write to a file (REALTIME) */
2113 #ifdef COMPAT_FREEBSD6
2114 int
2115 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2116 {
2117 
2118 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2119 	    &aiocb_ops_osigevent));
2120 }
2121 #endif
2122 
2123 int
2124 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2125 {
2126 
2127 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2128 }
2129 
2130 int
2131 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2132 {
2133 
2134 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2135 }
2136 
2137 static int
2138 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2139     struct aiocb **acb_list, int nent, struct sigevent *sig,
2140     struct aiocb_ops *ops)
2141 {
2142 	struct proc *p = td->td_proc;
2143 	struct aiocb *job;
2144 	struct kaioinfo *ki;
2145 	struct aioliojob *lj;
2146 	struct kevent kev;
2147 	int error;
2148 	int nerror;
2149 	int i;
2150 
2151 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2152 		return (EINVAL);
2153 
2154 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2155 		return (EINVAL);
2156 
2157 	if (p->p_aioinfo == NULL)
2158 		aio_init_aioinfo(p);
2159 
2160 	ki = p->p_aioinfo;
2161 
2162 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2163 	lj->lioj_flags = 0;
2164 	lj->lioj_count = 0;
2165 	lj->lioj_finished_count = 0;
2166 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2167 	ksiginfo_init(&lj->lioj_ksi);
2168 
2169 	/*
2170 	 * Setup signal.
2171 	 */
2172 	if (sig && (mode == LIO_NOWAIT)) {
2173 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2174 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2175 			/* Assume only new style KEVENT */
2176 			kev.filter = EVFILT_LIO;
2177 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2178 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2179 			kev.data = (intptr_t)lj;
2180 			/* pass user defined sigval data */
2181 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2182 			error = kqfd_register(
2183 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2184 			if (error) {
2185 				uma_zfree(aiolio_zone, lj);
2186 				return (error);
2187 			}
2188 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2189 			;
2190 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2191 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2192 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2193 					uma_zfree(aiolio_zone, lj);
2194 					return EINVAL;
2195 				}
2196 				lj->lioj_flags |= LIOJ_SIGNAL;
2197 		} else {
2198 			uma_zfree(aiolio_zone, lj);
2199 			return EINVAL;
2200 		}
2201 	}
2202 
2203 	AIO_LOCK(ki);
2204 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2205 	/*
2206 	 * Add extra aiocb count to avoid the lio to be freed
2207 	 * by other threads doing aio_waitcomplete or aio_return,
2208 	 * and prevent event from being sent until we have queued
2209 	 * all tasks.
2210 	 */
2211 	lj->lioj_count = 1;
2212 	AIO_UNLOCK(ki);
2213 
2214 	/*
2215 	 * Get pointers to the list of I/O requests.
2216 	 */
2217 	nerror = 0;
2218 	for (i = 0; i < nent; i++) {
2219 		job = acb_list[i];
2220 		if (job != NULL) {
2221 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2222 			if (error != 0)
2223 				nerror++;
2224 		}
2225 	}
2226 
2227 	error = 0;
2228 	AIO_LOCK(ki);
2229 	if (mode == LIO_WAIT) {
2230 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2231 			ki->kaio_flags |= KAIO_WAKEUP;
2232 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2233 			    PRIBIO | PCATCH, "aiospn", 0);
2234 			if (error == ERESTART)
2235 				error = EINTR;
2236 			if (error)
2237 				break;
2238 		}
2239 	} else {
2240 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2241 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2242 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2243 				KNOTE_LOCKED(&lj->klist, 1);
2244 			}
2245 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2246 			    == LIOJ_SIGNAL
2247 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2248 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2249 				aio_sendsig(p, &lj->lioj_signal,
2250 					    &lj->lioj_ksi);
2251 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2252 			}
2253 		}
2254 	}
2255 	lj->lioj_count--;
2256 	if (lj->lioj_count == 0) {
2257 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2258 		knlist_delete(&lj->klist, curthread, 1);
2259 		PROC_LOCK(p);
2260 		sigqueue_take(&lj->lioj_ksi);
2261 		PROC_UNLOCK(p);
2262 		AIO_UNLOCK(ki);
2263 		uma_zfree(aiolio_zone, lj);
2264 	} else
2265 		AIO_UNLOCK(ki);
2266 
2267 	if (nerror)
2268 		return (EIO);
2269 	return (error);
2270 }
2271 
2272 /* syscall - list directed I/O (REALTIME) */
2273 #ifdef COMPAT_FREEBSD6
2274 int
2275 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2276 {
2277 	struct aiocb **acb_list;
2278 	struct sigevent *sigp, sig;
2279 	struct osigevent osig;
2280 	int error, nent;
2281 
2282 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2283 		return (EINVAL);
2284 
2285 	nent = uap->nent;
2286 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2287 		return (EINVAL);
2288 
2289 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2290 		error = copyin(uap->sig, &osig, sizeof(osig));
2291 		if (error)
2292 			return (error);
2293 		error = convert_old_sigevent(&osig, &sig);
2294 		if (error)
2295 			return (error);
2296 		sigp = &sig;
2297 	} else
2298 		sigp = NULL;
2299 
2300 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2301 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2302 	if (error == 0)
2303 		error = kern_lio_listio(td, uap->mode,
2304 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2305 		    &aiocb_ops_osigevent);
2306 	free(acb_list, M_LIO);
2307 	return (error);
2308 }
2309 #endif
2310 
2311 /* syscall - list directed I/O (REALTIME) */
2312 int
2313 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2314 {
2315 	struct aiocb **acb_list;
2316 	struct sigevent *sigp, sig;
2317 	int error, nent;
2318 
2319 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2320 		return (EINVAL);
2321 
2322 	nent = uap->nent;
2323 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2324 		return (EINVAL);
2325 
2326 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2327 		error = copyin(uap->sig, &sig, sizeof(sig));
2328 		if (error)
2329 			return (error);
2330 		sigp = &sig;
2331 	} else
2332 		sigp = NULL;
2333 
2334 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2335 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2336 	if (error == 0)
2337 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2338 		    nent, sigp, &aiocb_ops);
2339 	free(acb_list, M_LIO);
2340 	return (error);
2341 }
2342 
2343 static void
2344 aio_physwakeup(struct bio *bp)
2345 {
2346 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2347 	struct proc *userp;
2348 	struct kaioinfo *ki;
2349 	size_t nbytes;
2350 	int error, nblks;
2351 
2352 	/* Release mapping into kernel space. */
2353 	userp = job->userproc;
2354 	ki = userp->p_aioinfo;
2355 	if (job->pbuf) {
2356 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2357 		relpbuf(job->pbuf, NULL);
2358 		job->pbuf = NULL;
2359 		atomic_subtract_int(&num_buf_aio, 1);
2360 		AIO_LOCK(ki);
2361 		ki->kaio_buffer_count--;
2362 		AIO_UNLOCK(ki);
2363 	}
2364 	vm_page_unhold_pages(job->pages, job->npages);
2365 
2366 	bp = job->bp;
2367 	job->bp = NULL;
2368 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2369 	error = 0;
2370 	if (bp->bio_flags & BIO_ERROR)
2371 		error = bp->bio_error;
2372 	nblks = btodb(nbytes);
2373 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2374 		job->outblock += nblks;
2375 	else
2376 		job->inblock += nblks;
2377 
2378 	if (error)
2379 		aio_complete(job, -1, error);
2380 	else
2381 		aio_complete(job, nbytes, 0);
2382 
2383 	g_destroy_bio(bp);
2384 }
2385 
2386 /* syscall - wait for the next completion of an aio request */
2387 static int
2388 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2389     struct timespec *ts, struct aiocb_ops *ops)
2390 {
2391 	struct proc *p = td->td_proc;
2392 	struct timeval atv;
2393 	struct kaioinfo *ki;
2394 	struct kaiocb *job;
2395 	struct aiocb *ujob;
2396 	long error, status;
2397 	int timo;
2398 
2399 	ops->store_aiocb(ujobp, NULL);
2400 
2401 	if (ts == NULL) {
2402 		timo = 0;
2403 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2404 		timo = -1;
2405 	} else {
2406 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2407 			return (EINVAL);
2408 
2409 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2410 		if (itimerfix(&atv))
2411 			return (EINVAL);
2412 		timo = tvtohz(&atv);
2413 	}
2414 
2415 	if (p->p_aioinfo == NULL)
2416 		aio_init_aioinfo(p);
2417 	ki = p->p_aioinfo;
2418 
2419 	error = 0;
2420 	job = NULL;
2421 	AIO_LOCK(ki);
2422 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2423 		if (timo == -1) {
2424 			error = EWOULDBLOCK;
2425 			break;
2426 		}
2427 		ki->kaio_flags |= KAIO_WAKEUP;
2428 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2429 		    "aiowc", timo);
2430 		if (timo && error == ERESTART)
2431 			error = EINTR;
2432 		if (error)
2433 			break;
2434 	}
2435 
2436 	if (job != NULL) {
2437 		MPASS(job->jobflags & KAIOCB_FINISHED);
2438 		ujob = job->ujob;
2439 		status = job->uaiocb._aiocb_private.status;
2440 		error = job->uaiocb._aiocb_private.error;
2441 		td->td_retval[0] = status;
2442 		td->td_ru.ru_oublock += job->outblock;
2443 		td->td_ru.ru_inblock += job->inblock;
2444 		td->td_ru.ru_msgsnd += job->msgsnd;
2445 		td->td_ru.ru_msgrcv += job->msgrcv;
2446 		aio_free_entry(job);
2447 		AIO_UNLOCK(ki);
2448 		ops->store_aiocb(ujobp, ujob);
2449 		ops->store_error(ujob, error);
2450 		ops->store_status(ujob, status);
2451 	} else
2452 		AIO_UNLOCK(ki);
2453 
2454 	return (error);
2455 }
2456 
2457 int
2458 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2459 {
2460 	struct timespec ts, *tsp;
2461 	int error;
2462 
2463 	if (uap->timeout) {
2464 		/* Get timespec struct. */
2465 		error = copyin(uap->timeout, &ts, sizeof(ts));
2466 		if (error)
2467 			return (error);
2468 		tsp = &ts;
2469 	} else
2470 		tsp = NULL;
2471 
2472 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2473 }
2474 
2475 static int
2476 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2477     struct aiocb_ops *ops)
2478 {
2479 
2480 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2481 		return (EINVAL);
2482 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2483 }
2484 
2485 int
2486 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2487 {
2488 
2489 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2490 }
2491 
2492 /* kqueue attach function */
2493 static int
2494 filt_aioattach(struct knote *kn)
2495 {
2496 	struct kaiocb *job = (struct kaiocb *)kn->kn_sdata;
2497 
2498 	/*
2499 	 * The job pointer must be validated before using it, so
2500 	 * registration is restricted to the kernel; the user cannot
2501 	 * set EV_FLAG1.
2502 	 */
2503 	if ((kn->kn_flags & EV_FLAG1) == 0)
2504 		return (EPERM);
2505 	kn->kn_ptr.p_aio = job;
2506 	kn->kn_flags &= ~EV_FLAG1;
2507 
2508 	knlist_add(&job->klist, kn, 0);
2509 
2510 	return (0);
2511 }
2512 
2513 /* kqueue detach function */
2514 static void
2515 filt_aiodetach(struct knote *kn)
2516 {
2517 	struct knlist *knl;
2518 
2519 	knl = &kn->kn_ptr.p_aio->klist;
2520 	knl->kl_lock(knl->kl_lockarg);
2521 	if (!knlist_empty(knl))
2522 		knlist_remove(knl, kn, 1);
2523 	knl->kl_unlock(knl->kl_lockarg);
2524 }
2525 
2526 /* kqueue filter function */
2527 /*ARGSUSED*/
2528 static int
2529 filt_aio(struct knote *kn, long hint)
2530 {
2531 	struct kaiocb *job = kn->kn_ptr.p_aio;
2532 
2533 	kn->kn_data = job->uaiocb._aiocb_private.error;
2534 	if (!(job->jobflags & KAIOCB_FINISHED))
2535 		return (0);
2536 	kn->kn_flags |= EV_EOF;
2537 	return (1);
2538 }
2539 
2540 /* kqueue attach function */
2541 static int
2542 filt_lioattach(struct knote *kn)
2543 {
2544 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2545 
2546 	/*
2547 	 * The aioliojob pointer must be validated before using it, so
2548 	 * registration is restricted to the kernel; the user cannot
2549 	 * set EV_FLAG1.
2550 	 */
2551 	if ((kn->kn_flags & EV_FLAG1) == 0)
2552 		return (EPERM);
2553 	kn->kn_ptr.p_lio = lj;
2554 	kn->kn_flags &= ~EV_FLAG1;
2555 
2556 	knlist_add(&lj->klist, kn, 0);
2557 
2558 	return (0);
2559 }
2560 
2561 /* kqueue detach function */
2562 static void
2563 filt_liodetach(struct knote *kn)
2564 {
2565 	struct knlist *knl;
2566 
2567 	knl = &kn->kn_ptr.p_lio->klist;
2568 	knl->kl_lock(knl->kl_lockarg);
2569 	if (!knlist_empty(knl))
2570 		knlist_remove(knl, kn, 1);
2571 	knl->kl_unlock(knl->kl_lockarg);
2572 }
2573 
2574 /* kqueue filter function */
2575 /*ARGSUSED*/
2576 static int
2577 filt_lio(struct knote *kn, long hint)
2578 {
2579 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2580 
2581 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2582 }
2583 
2584 #ifdef COMPAT_FREEBSD32
2585 #include <sys/mount.h>
2586 #include <sys/socket.h>
2587 #include <compat/freebsd32/freebsd32.h>
2588 #include <compat/freebsd32/freebsd32_proto.h>
2589 #include <compat/freebsd32/freebsd32_signal.h>
2590 #include <compat/freebsd32/freebsd32_syscall.h>
2591 #include <compat/freebsd32/freebsd32_util.h>
2592 
2593 struct __aiocb_private32 {
2594 	int32_t	status;
2595 	int32_t	error;
2596 	uint32_t kernelinfo;
2597 };
2598 
2599 #ifdef COMPAT_FREEBSD6
2600 typedef struct oaiocb32 {
2601 	int	aio_fildes;		/* File descriptor */
2602 	uint64_t aio_offset __packed;	/* File offset for I/O */
2603 	uint32_t aio_buf;		/* I/O buffer in process space */
2604 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2605 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2606 	int	aio_lio_opcode;		/* LIO opcode */
2607 	int	aio_reqprio;		/* Request priority -- ignored */
2608 	struct	__aiocb_private32 _aiocb_private;
2609 } oaiocb32_t;
2610 #endif
2611 
2612 typedef struct aiocb32 {
2613 	int32_t	aio_fildes;		/* File descriptor */
2614 	uint64_t aio_offset __packed;	/* File offset for I/O */
2615 	uint32_t aio_buf;		/* I/O buffer in process space */
2616 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2617 	int	__spare__[2];
2618 	uint32_t __spare2__;
2619 	int	aio_lio_opcode;		/* LIO opcode */
2620 	int	aio_reqprio;		/* Request priority -- ignored */
2621 	struct	__aiocb_private32 _aiocb_private;
2622 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2623 } aiocb32_t;
2624 
2625 #ifdef COMPAT_FREEBSD6
2626 static int
2627 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2628 {
2629 
2630 	/*
2631 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2632 	 * supported by AIO with the old sigevent structure.
2633 	 */
2634 	CP(*osig, *nsig, sigev_notify);
2635 	switch (nsig->sigev_notify) {
2636 	case SIGEV_NONE:
2637 		break;
2638 	case SIGEV_SIGNAL:
2639 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2640 		break;
2641 	case SIGEV_KEVENT:
2642 		nsig->sigev_notify_kqueue =
2643 		    osig->__sigev_u.__sigev_notify_kqueue;
2644 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2645 		break;
2646 	default:
2647 		return (EINVAL);
2648 	}
2649 	return (0);
2650 }
2651 
2652 static int
2653 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2654 {
2655 	struct oaiocb32 job32;
2656 	int error;
2657 
2658 	bzero(kjob, sizeof(struct aiocb));
2659 	error = copyin(ujob, &job32, sizeof(job32));
2660 	if (error)
2661 		return (error);
2662 
2663 	CP(job32, *kjob, aio_fildes);
2664 	CP(job32, *kjob, aio_offset);
2665 	PTRIN_CP(job32, *kjob, aio_buf);
2666 	CP(job32, *kjob, aio_nbytes);
2667 	CP(job32, *kjob, aio_lio_opcode);
2668 	CP(job32, *kjob, aio_reqprio);
2669 	CP(job32, *kjob, _aiocb_private.status);
2670 	CP(job32, *kjob, _aiocb_private.error);
2671 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2672 	return (convert_old_sigevent32(&job32.aio_sigevent,
2673 	    &kjob->aio_sigevent));
2674 }
2675 #endif
2676 
2677 static int
2678 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2679 {
2680 	struct aiocb32 job32;
2681 	int error;
2682 
2683 	error = copyin(ujob, &job32, sizeof(job32));
2684 	if (error)
2685 		return (error);
2686 	CP(job32, *kjob, aio_fildes);
2687 	CP(job32, *kjob, aio_offset);
2688 	PTRIN_CP(job32, *kjob, aio_buf);
2689 	CP(job32, *kjob, aio_nbytes);
2690 	CP(job32, *kjob, aio_lio_opcode);
2691 	CP(job32, *kjob, aio_reqprio);
2692 	CP(job32, *kjob, _aiocb_private.status);
2693 	CP(job32, *kjob, _aiocb_private.error);
2694 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2695 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2696 }
2697 
2698 static long
2699 aiocb32_fetch_status(struct aiocb *ujob)
2700 {
2701 	struct aiocb32 *ujob32;
2702 
2703 	ujob32 = (struct aiocb32 *)ujob;
2704 	return (fuword32(&ujob32->_aiocb_private.status));
2705 }
2706 
2707 static long
2708 aiocb32_fetch_error(struct aiocb *ujob)
2709 {
2710 	struct aiocb32 *ujob32;
2711 
2712 	ujob32 = (struct aiocb32 *)ujob;
2713 	return (fuword32(&ujob32->_aiocb_private.error));
2714 }
2715 
2716 static int
2717 aiocb32_store_status(struct aiocb *ujob, long status)
2718 {
2719 	struct aiocb32 *ujob32;
2720 
2721 	ujob32 = (struct aiocb32 *)ujob;
2722 	return (suword32(&ujob32->_aiocb_private.status, status));
2723 }
2724 
2725 static int
2726 aiocb32_store_error(struct aiocb *ujob, long error)
2727 {
2728 	struct aiocb32 *ujob32;
2729 
2730 	ujob32 = (struct aiocb32 *)ujob;
2731 	return (suword32(&ujob32->_aiocb_private.error, error));
2732 }
2733 
2734 static int
2735 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2736 {
2737 	struct aiocb32 *ujob32;
2738 
2739 	ujob32 = (struct aiocb32 *)ujob;
2740 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2741 }
2742 
2743 static int
2744 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2745 {
2746 
2747 	return (suword32(ujobp, (long)ujob));
2748 }
2749 
2750 static struct aiocb_ops aiocb32_ops = {
2751 	.copyin = aiocb32_copyin,
2752 	.fetch_status = aiocb32_fetch_status,
2753 	.fetch_error = aiocb32_fetch_error,
2754 	.store_status = aiocb32_store_status,
2755 	.store_error = aiocb32_store_error,
2756 	.store_kernelinfo = aiocb32_store_kernelinfo,
2757 	.store_aiocb = aiocb32_store_aiocb,
2758 };
2759 
2760 #ifdef COMPAT_FREEBSD6
2761 static struct aiocb_ops aiocb32_ops_osigevent = {
2762 	.copyin = aiocb32_copyin_old_sigevent,
2763 	.fetch_status = aiocb32_fetch_status,
2764 	.fetch_error = aiocb32_fetch_error,
2765 	.store_status = aiocb32_store_status,
2766 	.store_error = aiocb32_store_error,
2767 	.store_kernelinfo = aiocb32_store_kernelinfo,
2768 	.store_aiocb = aiocb32_store_aiocb,
2769 };
2770 #endif
2771 
2772 int
2773 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2774 {
2775 
2776 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2777 }
2778 
2779 int
2780 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2781 {
2782 	struct timespec32 ts32;
2783 	struct timespec ts, *tsp;
2784 	struct aiocb **ujoblist;
2785 	uint32_t *ujoblist32;
2786 	int error, i;
2787 
2788 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2789 		return (EINVAL);
2790 
2791 	if (uap->timeout) {
2792 		/* Get timespec struct. */
2793 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2794 			return (error);
2795 		CP(ts32, ts, tv_sec);
2796 		CP(ts32, ts, tv_nsec);
2797 		tsp = &ts;
2798 	} else
2799 		tsp = NULL;
2800 
2801 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2802 	ujoblist32 = (uint32_t *)ujoblist;
2803 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2804 	    sizeof(ujoblist32[0]));
2805 	if (error == 0) {
2806 		for (i = uap->nent; i > 0; i--)
2807 			ujoblist[i] = PTRIN(ujoblist32[i]);
2808 
2809 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2810 	}
2811 	uma_zfree(aiol_zone, ujoblist);
2812 	return (error);
2813 }
2814 
2815 int
2816 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2817 {
2818 
2819 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2820 }
2821 
2822 #ifdef COMPAT_FREEBSD6
2823 int
2824 freebsd6_freebsd32_aio_read(struct thread *td,
2825     struct freebsd6_freebsd32_aio_read_args *uap)
2826 {
2827 
2828 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2829 	    &aiocb32_ops_osigevent));
2830 }
2831 #endif
2832 
2833 int
2834 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2835 {
2836 
2837 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2838 	    &aiocb32_ops));
2839 }
2840 
2841 #ifdef COMPAT_FREEBSD6
2842 int
2843 freebsd6_freebsd32_aio_write(struct thread *td,
2844     struct freebsd6_freebsd32_aio_write_args *uap)
2845 {
2846 
2847 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2848 	    &aiocb32_ops_osigevent));
2849 }
2850 #endif
2851 
2852 int
2853 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2854 {
2855 
2856 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2857 	    &aiocb32_ops));
2858 }
2859 
2860 int
2861 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2862 {
2863 
2864 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2865 	    &aiocb32_ops));
2866 }
2867 
2868 int
2869 freebsd32_aio_waitcomplete(struct thread *td,
2870     struct freebsd32_aio_waitcomplete_args *uap)
2871 {
2872 	struct timespec32 ts32;
2873 	struct timespec ts, *tsp;
2874 	int error;
2875 
2876 	if (uap->timeout) {
2877 		/* Get timespec struct. */
2878 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2879 		if (error)
2880 			return (error);
2881 		CP(ts32, ts, tv_sec);
2882 		CP(ts32, ts, tv_nsec);
2883 		tsp = &ts;
2884 	} else
2885 		tsp = NULL;
2886 
2887 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2888 	    &aiocb32_ops));
2889 }
2890 
2891 int
2892 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2893 {
2894 
2895 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2896 	    &aiocb32_ops));
2897 }
2898 
2899 #ifdef COMPAT_FREEBSD6
2900 int
2901 freebsd6_freebsd32_lio_listio(struct thread *td,
2902     struct freebsd6_freebsd32_lio_listio_args *uap)
2903 {
2904 	struct aiocb **acb_list;
2905 	struct sigevent *sigp, sig;
2906 	struct osigevent32 osig;
2907 	uint32_t *acb_list32;
2908 	int error, i, nent;
2909 
2910 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2911 		return (EINVAL);
2912 
2913 	nent = uap->nent;
2914 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2915 		return (EINVAL);
2916 
2917 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2918 		error = copyin(uap->sig, &osig, sizeof(osig));
2919 		if (error)
2920 			return (error);
2921 		error = convert_old_sigevent32(&osig, &sig);
2922 		if (error)
2923 			return (error);
2924 		sigp = &sig;
2925 	} else
2926 		sigp = NULL;
2927 
2928 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2929 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2930 	if (error) {
2931 		free(acb_list32, M_LIO);
2932 		return (error);
2933 	}
2934 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2935 	for (i = 0; i < nent; i++)
2936 		acb_list[i] = PTRIN(acb_list32[i]);
2937 	free(acb_list32, M_LIO);
2938 
2939 	error = kern_lio_listio(td, uap->mode,
2940 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2941 	    &aiocb32_ops_osigevent);
2942 	free(acb_list, M_LIO);
2943 	return (error);
2944 }
2945 #endif
2946 
2947 int
2948 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2949 {
2950 	struct aiocb **acb_list;
2951 	struct sigevent *sigp, sig;
2952 	struct sigevent32 sig32;
2953 	uint32_t *acb_list32;
2954 	int error, i, nent;
2955 
2956 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2957 		return (EINVAL);
2958 
2959 	nent = uap->nent;
2960 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2961 		return (EINVAL);
2962 
2963 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2964 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2965 		if (error)
2966 			return (error);
2967 		error = convert_sigevent32(&sig32, &sig);
2968 		if (error)
2969 			return (error);
2970 		sigp = &sig;
2971 	} else
2972 		sigp = NULL;
2973 
2974 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2975 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2976 	if (error) {
2977 		free(acb_list32, M_LIO);
2978 		return (error);
2979 	}
2980 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2981 	for (i = 0; i < nent; i++)
2982 		acb_list[i] = PTRIN(acb_list32[i]);
2983 	free(acb_list32, M_LIO);
2984 
2985 	error = kern_lio_listio(td, uap->mode,
2986 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2987 	    &aiocb32_ops);
2988 	free(acb_list, M_LIO);
2989 	return (error);
2990 }
2991 
2992 #endif
2993