xref: /freebsd/sys/kern/vfs_aio.c (revision 0b57cec536236d46e3dba9bd041533462f33dbb7)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 1997 John S. Dyson.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. John S. Dyson's name may not be used to endorse or promote products
12  *    derived from this software without specific prior written permission.
13  *
14  * DISCLAIMER:  This code isn't warranted to do anything useful.  Anything
15  * bad that happens because of using this software isn't the responsibility
16  * of the author.  This software is distributed AS-IS.
17  */
18 
19 /*
20  * This file contains support for the POSIX 1003.1B AIO/LIO facility.
21  */
22 
23 #include <sys/cdefs.h>
24 __FBSDID("$FreeBSD$");
25 
26 #include <sys/param.h>
27 #include <sys/systm.h>
28 #include <sys/malloc.h>
29 #include <sys/bio.h>
30 #include <sys/buf.h>
31 #include <sys/capsicum.h>
32 #include <sys/eventhandler.h>
33 #include <sys/sysproto.h>
34 #include <sys/filedesc.h>
35 #include <sys/kernel.h>
36 #include <sys/module.h>
37 #include <sys/kthread.h>
38 #include <sys/fcntl.h>
39 #include <sys/file.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/mutex.h>
43 #include <sys/unistd.h>
44 #include <sys/posix4.h>
45 #include <sys/proc.h>
46 #include <sys/resourcevar.h>
47 #include <sys/signalvar.h>
48 #include <sys/syscallsubr.h>
49 #include <sys/protosw.h>
50 #include <sys/rwlock.h>
51 #include <sys/sema.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/syscall.h>
55 #include <sys/sysent.h>
56 #include <sys/sysctl.h>
57 #include <sys/syslog.h>
58 #include <sys/sx.h>
59 #include <sys/taskqueue.h>
60 #include <sys/vnode.h>
61 #include <sys/conf.h>
62 #include <sys/event.h>
63 #include <sys/mount.h>
64 #include <geom/geom.h>
65 
66 #include <machine/atomic.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_page.h>
70 #include <vm/vm_extern.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_map.h>
73 #include <vm/vm_object.h>
74 #include <vm/uma.h>
75 #include <sys/aio.h>
76 
77 /*
78  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
79  * overflow. (XXX will be removed soon.)
80  */
81 static u_long jobrefid;
82 
83 /*
84  * Counter for aio_fsync.
85  */
86 static uint64_t jobseqno;
87 
88 #ifndef MAX_AIO_PER_PROC
89 #define MAX_AIO_PER_PROC	32
90 #endif
91 
92 #ifndef MAX_AIO_QUEUE_PER_PROC
93 #define MAX_AIO_QUEUE_PER_PROC	256
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE
97 #define MAX_AIO_QUEUE		1024 /* Bigger than MAX_AIO_QUEUE_PER_PROC */
98 #endif
99 
100 #ifndef MAX_BUF_AIO
101 #define MAX_BUF_AIO		16
102 #endif
103 
104 FEATURE(aio, "Asynchronous I/O");
105 SYSCTL_DECL(_p1003_1b);
106 
107 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
108 static MALLOC_DEFINE(M_AIOS, "aios", "aio_suspend aio control block list");
109 
110 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0,
111     "Async IO management");
112 
113 static int enable_aio_unsafe = 0;
114 SYSCTL_INT(_vfs_aio, OID_AUTO, enable_unsafe, CTLFLAG_RW, &enable_aio_unsafe, 0,
115     "Permit asynchronous IO on all file types, not just known-safe types");
116 
117 static unsigned int unsafe_warningcnt = 1;
118 SYSCTL_UINT(_vfs_aio, OID_AUTO, unsafe_warningcnt, CTLFLAG_RW,
119     &unsafe_warningcnt, 0,
120     "Warnings that will be triggered upon failed IO requests on unsafe files");
121 
122 static int max_aio_procs = MAX_AIO_PROCS;
123 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs, CTLFLAG_RW, &max_aio_procs, 0,
124     "Maximum number of kernel processes to use for handling async IO ");
125 
126 static int num_aio_procs = 0;
127 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs, CTLFLAG_RD, &num_aio_procs, 0,
128     "Number of presently active kernel processes for async IO");
129 
130 /*
131  * The code will adjust the actual number of AIO processes towards this
132  * number when it gets a chance.
133  */
134 static int target_aio_procs = TARGET_AIO_PROCS;
135 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
136     0,
137     "Preferred number of ready kernel processes for async IO");
138 
139 static int max_queue_count = MAX_AIO_QUEUE;
140 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
141     "Maximum number of aio requests to queue, globally");
142 
143 static int num_queue_count = 0;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
145     "Number of queued aio requests");
146 
147 static int num_buf_aio = 0;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
149     "Number of aio requests presently handled by the buf subsystem");
150 
151 static int num_unmapped_aio = 0;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, num_unmapped_aio, CTLFLAG_RD, &num_unmapped_aio,
153     0,
154     "Number of aio requests presently handled by unmapped I/O buffers");
155 
156 /* Number of async I/O processes in the process of being started */
157 /* XXX This should be local to aio_aqueue() */
158 static int num_aio_resv_start = 0;
159 
160 static int aiod_lifetime;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
162     "Maximum lifetime for idle aiod");
163 
164 static int max_aio_per_proc = MAX_AIO_PER_PROC;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
166     0,
167     "Maximum active aio requests per process");
168 
169 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
170 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
171     &max_aio_queue_per_proc, 0,
172     "Maximum queued aio requests per process");
173 
174 static int max_buf_aio = MAX_BUF_AIO;
175 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
176     "Maximum buf aio requests per process");
177 
178 /*
179  * Though redundant with vfs.aio.max_aio_queue_per_proc, POSIX requires
180  * sysconf(3) to support AIO_LISTIO_MAX, and we implement that with
181  * vfs.aio.aio_listio_max.
182  */
183 SYSCTL_INT(_p1003_1b, CTL_P1003_1B_AIO_LISTIO_MAX, aio_listio_max,
184     CTLFLAG_RD | CTLFLAG_CAPRD, &max_aio_queue_per_proc,
185     0, "Maximum aio requests for a single lio_listio call");
186 
187 #ifdef COMPAT_FREEBSD6
188 typedef struct oaiocb {
189 	int	aio_fildes;		/* File descriptor */
190 	off_t	aio_offset;		/* File offset for I/O */
191 	volatile void *aio_buf;         /* I/O buffer in process space */
192 	size_t	aio_nbytes;		/* Number of bytes for I/O */
193 	struct	osigevent aio_sigevent;	/* Signal to deliver */
194 	int	aio_lio_opcode;		/* LIO opcode */
195 	int	aio_reqprio;		/* Request priority -- ignored */
196 	struct	__aiocb_private	_aiocb_private;
197 } oaiocb_t;
198 #endif
199 
200 /*
201  * Below is a key of locks used to protect each member of struct kaiocb
202  * aioliojob and kaioinfo and any backends.
203  *
204  * * - need not protected
205  * a - locked by kaioinfo lock
206  * b - locked by backend lock, the backend lock can be null in some cases,
207  *     for example, BIO belongs to this type, in this case, proc lock is
208  *     reused.
209  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
210  */
211 
212 /*
213  * If the routine that services an AIO request blocks while running in an
214  * AIO kernel process it can starve other I/O requests.  BIO requests
215  * queued via aio_qbio() complete asynchronously and do not use AIO kernel
216  * processes at all.  Socket I/O requests use a separate pool of
217  * kprocs and also force non-blocking I/O.  Other file I/O requests
218  * use the generic fo_read/fo_write operations which can block.  The
219  * fsync and mlock operations can also block while executing.  Ideally
220  * none of these requests would block while executing.
221  *
222  * Note that the service routines cannot toggle O_NONBLOCK in the file
223  * structure directly while handling a request due to races with
224  * userland threads.
225  */
226 
227 /* jobflags */
228 #define	KAIOCB_QUEUEING		0x01
229 #define	KAIOCB_CANCELLED	0x02
230 #define	KAIOCB_CANCELLING	0x04
231 #define	KAIOCB_CHECKSYNC	0x08
232 #define	KAIOCB_CLEARED		0x10
233 #define	KAIOCB_FINISHED		0x20
234 
235 /*
236  * AIO process info
237  */
238 #define AIOP_FREE	0x1			/* proc on free queue */
239 
240 struct aioproc {
241 	int	aioprocflags;			/* (c) AIO proc flags */
242 	TAILQ_ENTRY(aioproc) list;		/* (c) list of processes */
243 	struct	proc *aioproc;			/* (*) the AIO proc */
244 };
245 
246 /*
247  * data-structure for lio signal management
248  */
249 struct aioliojob {
250 	int	lioj_flags;			/* (a) listio flags */
251 	int	lioj_count;			/* (a) listio flags */
252 	int	lioj_finished_count;		/* (a) listio flags */
253 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
254 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
255 	struct	knlist klist;			/* (a) list of knotes */
256 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
257 };
258 
259 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
260 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
261 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
262 
263 /*
264  * per process aio data structure
265  */
266 struct kaioinfo {
267 	struct	mtx kaio_mtx;		/* the lock to protect this struct */
268 	int	kaio_flags;		/* (a) per process kaio flags */
269 	int	kaio_active_count;	/* (c) number of currently used AIOs */
270 	int	kaio_count;		/* (a) size of AIO queue */
271 	int	kaio_buffer_count;	/* (a) number of bio buffers */
272 	TAILQ_HEAD(,kaiocb) kaio_all;	/* (a) all AIOs in a process */
273 	TAILQ_HEAD(,kaiocb) kaio_done;	/* (a) done queue for process */
274 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
275 	TAILQ_HEAD(,kaiocb) kaio_jobqueue;	/* (a) job queue for process */
276 	TAILQ_HEAD(,kaiocb) kaio_syncqueue;	/* (a) queue for aio_fsync */
277 	TAILQ_HEAD(,kaiocb) kaio_syncready;  /* (a) second q for aio_fsync */
278 	struct	task kaio_task;		/* (*) task to kick aio processes */
279 	struct	task kaio_sync_task;	/* (*) task to schedule fsync jobs */
280 };
281 
282 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
283 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
284 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
285 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
286 
287 #define KAIO_RUNDOWN	0x1	/* process is being run down */
288 #define KAIO_WAKEUP	0x2	/* wakeup process when AIO completes */
289 
290 /*
291  * Operations used to interact with userland aio control blocks.
292  * Different ABIs provide their own operations.
293  */
294 struct aiocb_ops {
295 	int	(*aio_copyin)(struct aiocb *ujob, struct aiocb *kjob);
296 	long	(*fetch_status)(struct aiocb *ujob);
297 	long	(*fetch_error)(struct aiocb *ujob);
298 	int	(*store_status)(struct aiocb *ujob, long status);
299 	int	(*store_error)(struct aiocb *ujob, long error);
300 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
301 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
302 };
303 
304 static TAILQ_HEAD(,aioproc) aio_freeproc;		/* (c) Idle daemons */
305 static struct sema aio_newproc_sem;
306 static struct mtx aio_job_mtx;
307 static TAILQ_HEAD(,kaiocb) aio_jobs;			/* (c) Async job list */
308 static struct unrhdr *aiod_unr;
309 
310 void		aio_init_aioinfo(struct proc *p);
311 static int	aio_onceonly(void);
312 static int	aio_free_entry(struct kaiocb *job);
313 static void	aio_process_rw(struct kaiocb *job);
314 static void	aio_process_sync(struct kaiocb *job);
315 static void	aio_process_mlock(struct kaiocb *job);
316 static void	aio_schedule_fsync(void *context, int pending);
317 static int	aio_newproc(int *);
318 int		aio_aqueue(struct thread *td, struct aiocb *ujob,
319 		    struct aioliojob *lio, int type, struct aiocb_ops *ops);
320 static int	aio_queue_file(struct file *fp, struct kaiocb *job);
321 static void	aio_biowakeup(struct bio *bp);
322 static void	aio_proc_rundown(void *arg, struct proc *p);
323 static void	aio_proc_rundown_exec(void *arg, struct proc *p,
324 		    struct image_params *imgp);
325 static int	aio_qbio(struct proc *p, struct kaiocb *job);
326 static void	aio_daemon(void *param);
327 static void	aio_bio_done_notify(struct proc *userp, struct kaiocb *job);
328 static bool	aio_clear_cancel_function_locked(struct kaiocb *job);
329 static int	aio_kick(struct proc *userp);
330 static void	aio_kick_nowait(struct proc *userp);
331 static void	aio_kick_helper(void *context, int pending);
332 static int	filt_aioattach(struct knote *kn);
333 static void	filt_aiodetach(struct knote *kn);
334 static int	filt_aio(struct knote *kn, long hint);
335 static int	filt_lioattach(struct knote *kn);
336 static void	filt_liodetach(struct knote *kn);
337 static int	filt_lio(struct knote *kn, long hint);
338 
339 /*
340  * Zones for:
341  * 	kaio	Per process async io info
342  *	aiop	async io process data
343  *	aiocb	async io jobs
344  *	aiolio	list io jobs
345  */
346 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiolio_zone;
347 
348 /* kqueue filters for aio */
349 static struct filterops aio_filtops = {
350 	.f_isfd = 0,
351 	.f_attach = filt_aioattach,
352 	.f_detach = filt_aiodetach,
353 	.f_event = filt_aio,
354 };
355 static struct filterops lio_filtops = {
356 	.f_isfd = 0,
357 	.f_attach = filt_lioattach,
358 	.f_detach = filt_liodetach,
359 	.f_event = filt_lio
360 };
361 
362 static eventhandler_tag exit_tag, exec_tag;
363 
364 TASKQUEUE_DEFINE_THREAD(aiod_kick);
365 
366 /*
367  * Main operations function for use as a kernel module.
368  */
369 static int
370 aio_modload(struct module *module, int cmd, void *arg)
371 {
372 	int error = 0;
373 
374 	switch (cmd) {
375 	case MOD_LOAD:
376 		aio_onceonly();
377 		break;
378 	case MOD_SHUTDOWN:
379 		break;
380 	default:
381 		error = EOPNOTSUPP;
382 		break;
383 	}
384 	return (error);
385 }
386 
387 static moduledata_t aio_mod = {
388 	"aio",
389 	&aio_modload,
390 	NULL
391 };
392 
393 DECLARE_MODULE(aio, aio_mod, SI_SUB_VFS, SI_ORDER_ANY);
394 MODULE_VERSION(aio, 1);
395 
396 /*
397  * Startup initialization
398  */
399 static int
400 aio_onceonly(void)
401 {
402 
403 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
404 	    EVENTHANDLER_PRI_ANY);
405 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec,
406 	    NULL, EVENTHANDLER_PRI_ANY);
407 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
408 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
409 	TAILQ_INIT(&aio_freeproc);
410 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
411 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
412 	TAILQ_INIT(&aio_jobs);
413 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
414 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
415 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
416 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aioproc), NULL,
417 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
418 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct kaiocb), NULL, NULL,
419 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
420 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
421 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
422 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
423 	jobrefid = 1;
424 	p31b_setcfg(CTL_P1003_1B_ASYNCHRONOUS_IO, _POSIX_ASYNCHRONOUS_IO);
425 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
426 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
427 
428 	return (0);
429 }
430 
431 /*
432  * Init the per-process aioinfo structure.  The aioinfo limits are set
433  * per-process for user limit (resource) management.
434  */
435 void
436 aio_init_aioinfo(struct proc *p)
437 {
438 	struct kaioinfo *ki;
439 
440 	ki = uma_zalloc(kaio_zone, M_WAITOK);
441 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF | MTX_NEW);
442 	ki->kaio_flags = 0;
443 	ki->kaio_active_count = 0;
444 	ki->kaio_count = 0;
445 	ki->kaio_buffer_count = 0;
446 	TAILQ_INIT(&ki->kaio_all);
447 	TAILQ_INIT(&ki->kaio_done);
448 	TAILQ_INIT(&ki->kaio_jobqueue);
449 	TAILQ_INIT(&ki->kaio_liojoblist);
450 	TAILQ_INIT(&ki->kaio_syncqueue);
451 	TAILQ_INIT(&ki->kaio_syncready);
452 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
453 	TASK_INIT(&ki->kaio_sync_task, 0, aio_schedule_fsync, ki);
454 	PROC_LOCK(p);
455 	if (p->p_aioinfo == NULL) {
456 		p->p_aioinfo = ki;
457 		PROC_UNLOCK(p);
458 	} else {
459 		PROC_UNLOCK(p);
460 		mtx_destroy(&ki->kaio_mtx);
461 		uma_zfree(kaio_zone, ki);
462 	}
463 
464 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
465 		aio_newproc(NULL);
466 }
467 
468 static int
469 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
470 {
471 	struct thread *td;
472 	int error;
473 
474 	error = sigev_findtd(p, sigev, &td);
475 	if (error)
476 		return (error);
477 	if (!KSI_ONQ(ksi)) {
478 		ksiginfo_set_sigev(ksi, sigev);
479 		ksi->ksi_code = SI_ASYNCIO;
480 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
481 		tdsendsignal(p, td, ksi->ksi_signo, ksi);
482 	}
483 	PROC_UNLOCK(p);
484 	return (error);
485 }
486 
487 /*
488  * Free a job entry.  Wait for completion if it is currently active, but don't
489  * delay forever.  If we delay, we return a flag that says that we have to
490  * restart the queue scan.
491  */
492 static int
493 aio_free_entry(struct kaiocb *job)
494 {
495 	struct kaioinfo *ki;
496 	struct aioliojob *lj;
497 	struct proc *p;
498 
499 	p = job->userproc;
500 	MPASS(curproc == p);
501 	ki = p->p_aioinfo;
502 	MPASS(ki != NULL);
503 
504 	AIO_LOCK_ASSERT(ki, MA_OWNED);
505 	MPASS(job->jobflags & KAIOCB_FINISHED);
506 
507 	atomic_subtract_int(&num_queue_count, 1);
508 
509 	ki->kaio_count--;
510 	MPASS(ki->kaio_count >= 0);
511 
512 	TAILQ_REMOVE(&ki->kaio_done, job, plist);
513 	TAILQ_REMOVE(&ki->kaio_all, job, allist);
514 
515 	lj = job->lio;
516 	if (lj) {
517 		lj->lioj_count--;
518 		lj->lioj_finished_count--;
519 
520 		if (lj->lioj_count == 0) {
521 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
522 			/* lio is going away, we need to destroy any knotes */
523 			knlist_delete(&lj->klist, curthread, 1);
524 			PROC_LOCK(p);
525 			sigqueue_take(&lj->lioj_ksi);
526 			PROC_UNLOCK(p);
527 			uma_zfree(aiolio_zone, lj);
528 		}
529 	}
530 
531 	/* job is going away, we need to destroy any knotes */
532 	knlist_delete(&job->klist, curthread, 1);
533 	PROC_LOCK(p);
534 	sigqueue_take(&job->ksi);
535 	PROC_UNLOCK(p);
536 
537 	AIO_UNLOCK(ki);
538 
539 	/*
540 	 * The thread argument here is used to find the owning process
541 	 * and is also passed to fo_close() which may pass it to various
542 	 * places such as devsw close() routines.  Because of that, we
543 	 * need a thread pointer from the process owning the job that is
544 	 * persistent and won't disappear out from under us or move to
545 	 * another process.
546 	 *
547 	 * Currently, all the callers of this function call it to remove
548 	 * a kaiocb from the current process' job list either via a
549 	 * syscall or due to the current process calling exit() or
550 	 * execve().  Thus, we know that p == curproc.  We also know that
551 	 * curthread can't exit since we are curthread.
552 	 *
553 	 * Therefore, we use curthread as the thread to pass to
554 	 * knlist_delete().  This does mean that it is possible for the
555 	 * thread pointer at close time to differ from the thread pointer
556 	 * at open time, but this is already true of file descriptors in
557 	 * a multithreaded process.
558 	 */
559 	if (job->fd_file)
560 		fdrop(job->fd_file, curthread);
561 	crfree(job->cred);
562 	uma_zfree(aiocb_zone, job);
563 	AIO_LOCK(ki);
564 
565 	return (0);
566 }
567 
568 static void
569 aio_proc_rundown_exec(void *arg, struct proc *p,
570     struct image_params *imgp __unused)
571 {
572    	aio_proc_rundown(arg, p);
573 }
574 
575 static int
576 aio_cancel_job(struct proc *p, struct kaioinfo *ki, struct kaiocb *job)
577 {
578 	aio_cancel_fn_t *func;
579 	int cancelled;
580 
581 	AIO_LOCK_ASSERT(ki, MA_OWNED);
582 	if (job->jobflags & (KAIOCB_CANCELLED | KAIOCB_FINISHED))
583 		return (0);
584 	MPASS((job->jobflags & KAIOCB_CANCELLING) == 0);
585 	job->jobflags |= KAIOCB_CANCELLED;
586 
587 	func = job->cancel_fn;
588 
589 	/*
590 	 * If there is no cancel routine, just leave the job marked as
591 	 * cancelled.  The job should be in active use by a caller who
592 	 * should complete it normally or when it fails to install a
593 	 * cancel routine.
594 	 */
595 	if (func == NULL)
596 		return (0);
597 
598 	/*
599 	 * Set the CANCELLING flag so that aio_complete() will defer
600 	 * completions of this job.  This prevents the job from being
601 	 * freed out from under the cancel callback.  After the
602 	 * callback any deferred completion (whether from the callback
603 	 * or any other source) will be completed.
604 	 */
605 	job->jobflags |= KAIOCB_CANCELLING;
606 	AIO_UNLOCK(ki);
607 	func(job);
608 	AIO_LOCK(ki);
609 	job->jobflags &= ~KAIOCB_CANCELLING;
610 	if (job->jobflags & KAIOCB_FINISHED) {
611 		cancelled = job->uaiocb._aiocb_private.error == ECANCELED;
612 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
613 		aio_bio_done_notify(p, job);
614 	} else {
615 		/*
616 		 * The cancel callback might have scheduled an
617 		 * operation to cancel this request, but it is
618 		 * only counted as cancelled if the request is
619 		 * cancelled when the callback returns.
620 		 */
621 		cancelled = 0;
622 	}
623 	return (cancelled);
624 }
625 
626 /*
627  * Rundown the jobs for a given process.
628  */
629 static void
630 aio_proc_rundown(void *arg, struct proc *p)
631 {
632 	struct kaioinfo *ki;
633 	struct aioliojob *lj;
634 	struct kaiocb *job, *jobn;
635 
636 	KASSERT(curthread->td_proc == p,
637 	    ("%s: called on non-curproc", __func__));
638 	ki = p->p_aioinfo;
639 	if (ki == NULL)
640 		return;
641 
642 	AIO_LOCK(ki);
643 	ki->kaio_flags |= KAIO_RUNDOWN;
644 
645 restart:
646 
647 	/*
648 	 * Try to cancel all pending requests. This code simulates
649 	 * aio_cancel on all pending I/O requests.
650 	 */
651 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
652 		aio_cancel_job(p, ki, job);
653 	}
654 
655 	/* Wait for all running I/O to be finished */
656 	if (TAILQ_FIRST(&ki->kaio_jobqueue) || ki->kaio_active_count != 0) {
657 		ki->kaio_flags |= KAIO_WAKEUP;
658 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
659 		goto restart;
660 	}
661 
662 	/* Free all completed I/O requests. */
663 	while ((job = TAILQ_FIRST(&ki->kaio_done)) != NULL)
664 		aio_free_entry(job);
665 
666 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
667 		if (lj->lioj_count == 0) {
668 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
669 			knlist_delete(&lj->klist, curthread, 1);
670 			PROC_LOCK(p);
671 			sigqueue_take(&lj->lioj_ksi);
672 			PROC_UNLOCK(p);
673 			uma_zfree(aiolio_zone, lj);
674 		} else {
675 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
676 			    lj->lioj_count, lj->lioj_finished_count);
677 		}
678 	}
679 	AIO_UNLOCK(ki);
680 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_task);
681 	taskqueue_drain(taskqueue_aiod_kick, &ki->kaio_sync_task);
682 	mtx_destroy(&ki->kaio_mtx);
683 	uma_zfree(kaio_zone, ki);
684 	p->p_aioinfo = NULL;
685 }
686 
687 /*
688  * Select a job to run (called by an AIO daemon).
689  */
690 static struct kaiocb *
691 aio_selectjob(struct aioproc *aiop)
692 {
693 	struct kaiocb *job;
694 	struct kaioinfo *ki;
695 	struct proc *userp;
696 
697 	mtx_assert(&aio_job_mtx, MA_OWNED);
698 restart:
699 	TAILQ_FOREACH(job, &aio_jobs, list) {
700 		userp = job->userproc;
701 		ki = userp->p_aioinfo;
702 
703 		if (ki->kaio_active_count < max_aio_per_proc) {
704 			TAILQ_REMOVE(&aio_jobs, job, list);
705 			if (!aio_clear_cancel_function(job))
706 				goto restart;
707 
708 			/* Account for currently active jobs. */
709 			ki->kaio_active_count++;
710 			break;
711 		}
712 	}
713 	return (job);
714 }
715 
716 /*
717  * Move all data to a permanent storage device.  This code
718  * simulates the fsync syscall.
719  */
720 static int
721 aio_fsync_vnode(struct thread *td, struct vnode *vp)
722 {
723 	struct mount *mp;
724 	int error;
725 
726 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
727 		goto drop;
728 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
729 	if (vp->v_object != NULL) {
730 		VM_OBJECT_WLOCK(vp->v_object);
731 		vm_object_page_clean(vp->v_object, 0, 0, 0);
732 		VM_OBJECT_WUNLOCK(vp->v_object);
733 	}
734 	error = VOP_FSYNC(vp, MNT_WAIT, td);
735 
736 	VOP_UNLOCK(vp, 0);
737 	vn_finished_write(mp);
738 drop:
739 	return (error);
740 }
741 
742 /*
743  * The AIO processing activity for LIO_READ/LIO_WRITE.  This is the code that
744  * does the I/O request for the non-bio version of the operations.  The normal
745  * vn operations are used, and this code should work in all instances for every
746  * type of file, including pipes, sockets, fifos, and regular files.
747  *
748  * XXX I don't think it works well for socket, pipe, and fifo.
749  */
750 static void
751 aio_process_rw(struct kaiocb *job)
752 {
753 	struct ucred *td_savedcred;
754 	struct thread *td;
755 	struct aiocb *cb;
756 	struct file *fp;
757 	struct uio auio;
758 	struct iovec aiov;
759 	ssize_t cnt;
760 	long msgsnd_st, msgsnd_end;
761 	long msgrcv_st, msgrcv_end;
762 	long oublock_st, oublock_end;
763 	long inblock_st, inblock_end;
764 	int error;
765 
766 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_READ ||
767 	    job->uaiocb.aio_lio_opcode == LIO_WRITE,
768 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
769 
770 	aio_switch_vmspace(job);
771 	td = curthread;
772 	td_savedcred = td->td_ucred;
773 	td->td_ucred = job->cred;
774 	cb = &job->uaiocb;
775 	fp = job->fd_file;
776 
777 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
778 	aiov.iov_len = cb->aio_nbytes;
779 
780 	auio.uio_iov = &aiov;
781 	auio.uio_iovcnt = 1;
782 	auio.uio_offset = cb->aio_offset;
783 	auio.uio_resid = cb->aio_nbytes;
784 	cnt = cb->aio_nbytes;
785 	auio.uio_segflg = UIO_USERSPACE;
786 	auio.uio_td = td;
787 
788 	msgrcv_st = td->td_ru.ru_msgrcv;
789 	msgsnd_st = td->td_ru.ru_msgsnd;
790 	inblock_st = td->td_ru.ru_inblock;
791 	oublock_st = td->td_ru.ru_oublock;
792 
793 	/*
794 	 * aio_aqueue() acquires a reference to the file that is
795 	 * released in aio_free_entry().
796 	 */
797 	if (cb->aio_lio_opcode == LIO_READ) {
798 		auio.uio_rw = UIO_READ;
799 		if (auio.uio_resid == 0)
800 			error = 0;
801 		else
802 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
803 	} else {
804 		if (fp->f_type == DTYPE_VNODE)
805 			bwillwrite();
806 		auio.uio_rw = UIO_WRITE;
807 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
808 	}
809 	msgrcv_end = td->td_ru.ru_msgrcv;
810 	msgsnd_end = td->td_ru.ru_msgsnd;
811 	inblock_end = td->td_ru.ru_inblock;
812 	oublock_end = td->td_ru.ru_oublock;
813 
814 	job->msgrcv = msgrcv_end - msgrcv_st;
815 	job->msgsnd = msgsnd_end - msgsnd_st;
816 	job->inblock = inblock_end - inblock_st;
817 	job->outblock = oublock_end - oublock_st;
818 
819 	if ((error) && (auio.uio_resid != cnt)) {
820 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
821 			error = 0;
822 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
823 			PROC_LOCK(job->userproc);
824 			kern_psignal(job->userproc, SIGPIPE);
825 			PROC_UNLOCK(job->userproc);
826 		}
827 	}
828 
829 	cnt -= auio.uio_resid;
830 	td->td_ucred = td_savedcred;
831 	if (error)
832 		aio_complete(job, -1, error);
833 	else
834 		aio_complete(job, cnt, 0);
835 }
836 
837 static void
838 aio_process_sync(struct kaiocb *job)
839 {
840 	struct thread *td = curthread;
841 	struct ucred *td_savedcred = td->td_ucred;
842 	struct file *fp = job->fd_file;
843 	int error = 0;
844 
845 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_SYNC,
846 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
847 
848 	td->td_ucred = job->cred;
849 	if (fp->f_vnode != NULL)
850 		error = aio_fsync_vnode(td, fp->f_vnode);
851 	td->td_ucred = td_savedcred;
852 	if (error)
853 		aio_complete(job, -1, error);
854 	else
855 		aio_complete(job, 0, 0);
856 }
857 
858 static void
859 aio_process_mlock(struct kaiocb *job)
860 {
861 	struct aiocb *cb = &job->uaiocb;
862 	int error;
863 
864 	KASSERT(job->uaiocb.aio_lio_opcode == LIO_MLOCK,
865 	    ("%s: opcode %d", __func__, job->uaiocb.aio_lio_opcode));
866 
867 	aio_switch_vmspace(job);
868 	error = kern_mlock(job->userproc, job->cred,
869 	    __DEVOLATILE(uintptr_t, cb->aio_buf), cb->aio_nbytes);
870 	aio_complete(job, error != 0 ? -1 : 0, error);
871 }
872 
873 static void
874 aio_bio_done_notify(struct proc *userp, struct kaiocb *job)
875 {
876 	struct aioliojob *lj;
877 	struct kaioinfo *ki;
878 	struct kaiocb *sjob, *sjobn;
879 	int lj_done;
880 	bool schedule_fsync;
881 
882 	ki = userp->p_aioinfo;
883 	AIO_LOCK_ASSERT(ki, MA_OWNED);
884 	lj = job->lio;
885 	lj_done = 0;
886 	if (lj) {
887 		lj->lioj_finished_count++;
888 		if (lj->lioj_count == lj->lioj_finished_count)
889 			lj_done = 1;
890 	}
891 	TAILQ_INSERT_TAIL(&ki->kaio_done, job, plist);
892 	MPASS(job->jobflags & KAIOCB_FINISHED);
893 
894 	if (ki->kaio_flags & KAIO_RUNDOWN)
895 		goto notification_done;
896 
897 	if (job->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
898 	    job->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
899 		aio_sendsig(userp, &job->uaiocb.aio_sigevent, &job->ksi);
900 
901 	KNOTE_LOCKED(&job->klist, 1);
902 
903 	if (lj_done) {
904 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
905 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
906 			KNOTE_LOCKED(&lj->klist, 1);
907 		}
908 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
909 		    == LIOJ_SIGNAL
910 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
911 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
912 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
913 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
914 		}
915 	}
916 
917 notification_done:
918 	if (job->jobflags & KAIOCB_CHECKSYNC) {
919 		schedule_fsync = false;
920 		TAILQ_FOREACH_SAFE(sjob, &ki->kaio_syncqueue, list, sjobn) {
921 			if (job->fd_file != sjob->fd_file ||
922 			    job->seqno >= sjob->seqno)
923 				continue;
924 			if (--sjob->pending > 0)
925 				continue;
926 			TAILQ_REMOVE(&ki->kaio_syncqueue, sjob, list);
927 			if (!aio_clear_cancel_function_locked(sjob))
928 				continue;
929 			TAILQ_INSERT_TAIL(&ki->kaio_syncready, sjob, list);
930 			schedule_fsync = true;
931 		}
932 		if (schedule_fsync)
933 			taskqueue_enqueue(taskqueue_aiod_kick,
934 			    &ki->kaio_sync_task);
935 	}
936 	if (ki->kaio_flags & KAIO_WAKEUP) {
937 		ki->kaio_flags &= ~KAIO_WAKEUP;
938 		wakeup(&userp->p_aioinfo);
939 	}
940 }
941 
942 static void
943 aio_schedule_fsync(void *context, int pending)
944 {
945 	struct kaioinfo *ki;
946 	struct kaiocb *job;
947 
948 	ki = context;
949 	AIO_LOCK(ki);
950 	while (!TAILQ_EMPTY(&ki->kaio_syncready)) {
951 		job = TAILQ_FIRST(&ki->kaio_syncready);
952 		TAILQ_REMOVE(&ki->kaio_syncready, job, list);
953 		AIO_UNLOCK(ki);
954 		aio_schedule(job, aio_process_sync);
955 		AIO_LOCK(ki);
956 	}
957 	AIO_UNLOCK(ki);
958 }
959 
960 bool
961 aio_cancel_cleared(struct kaiocb *job)
962 {
963 
964 	/*
965 	 * The caller should hold the same queue lock held when
966 	 * aio_clear_cancel_function() was called and set this flag
967 	 * ensuring this check sees an up-to-date value.  However,
968 	 * there is no way to assert that.
969 	 */
970 	return ((job->jobflags & KAIOCB_CLEARED) != 0);
971 }
972 
973 static bool
974 aio_clear_cancel_function_locked(struct kaiocb *job)
975 {
976 
977 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
978 	MPASS(job->cancel_fn != NULL);
979 	if (job->jobflags & KAIOCB_CANCELLING) {
980 		job->jobflags |= KAIOCB_CLEARED;
981 		return (false);
982 	}
983 	job->cancel_fn = NULL;
984 	return (true);
985 }
986 
987 bool
988 aio_clear_cancel_function(struct kaiocb *job)
989 {
990 	struct kaioinfo *ki;
991 	bool ret;
992 
993 	ki = job->userproc->p_aioinfo;
994 	AIO_LOCK(ki);
995 	ret = aio_clear_cancel_function_locked(job);
996 	AIO_UNLOCK(ki);
997 	return (ret);
998 }
999 
1000 static bool
1001 aio_set_cancel_function_locked(struct kaiocb *job, aio_cancel_fn_t *func)
1002 {
1003 
1004 	AIO_LOCK_ASSERT(job->userproc->p_aioinfo, MA_OWNED);
1005 	if (job->jobflags & KAIOCB_CANCELLED)
1006 		return (false);
1007 	job->cancel_fn = func;
1008 	return (true);
1009 }
1010 
1011 bool
1012 aio_set_cancel_function(struct kaiocb *job, aio_cancel_fn_t *func)
1013 {
1014 	struct kaioinfo *ki;
1015 	bool ret;
1016 
1017 	ki = job->userproc->p_aioinfo;
1018 	AIO_LOCK(ki);
1019 	ret = aio_set_cancel_function_locked(job, func);
1020 	AIO_UNLOCK(ki);
1021 	return (ret);
1022 }
1023 
1024 void
1025 aio_complete(struct kaiocb *job, long status, int error)
1026 {
1027 	struct kaioinfo *ki;
1028 	struct proc *userp;
1029 
1030 	job->uaiocb._aiocb_private.error = error;
1031 	job->uaiocb._aiocb_private.status = status;
1032 
1033 	userp = job->userproc;
1034 	ki = userp->p_aioinfo;
1035 
1036 	AIO_LOCK(ki);
1037 	KASSERT(!(job->jobflags & KAIOCB_FINISHED),
1038 	    ("duplicate aio_complete"));
1039 	job->jobflags |= KAIOCB_FINISHED;
1040 	if ((job->jobflags & (KAIOCB_QUEUEING | KAIOCB_CANCELLING)) == 0) {
1041 		TAILQ_REMOVE(&ki->kaio_jobqueue, job, plist);
1042 		aio_bio_done_notify(userp, job);
1043 	}
1044 	AIO_UNLOCK(ki);
1045 }
1046 
1047 void
1048 aio_cancel(struct kaiocb *job)
1049 {
1050 
1051 	aio_complete(job, -1, ECANCELED);
1052 }
1053 
1054 void
1055 aio_switch_vmspace(struct kaiocb *job)
1056 {
1057 
1058 	vmspace_switch_aio(job->userproc->p_vmspace);
1059 }
1060 
1061 /*
1062  * The AIO daemon, most of the actual work is done in aio_process_*,
1063  * but the setup (and address space mgmt) is done in this routine.
1064  */
1065 static void
1066 aio_daemon(void *_id)
1067 {
1068 	struct kaiocb *job;
1069 	struct aioproc *aiop;
1070 	struct kaioinfo *ki;
1071 	struct proc *p;
1072 	struct vmspace *myvm;
1073 	struct thread *td = curthread;
1074 	int id = (intptr_t)_id;
1075 
1076 	/*
1077 	 * Grab an extra reference on the daemon's vmspace so that it
1078 	 * doesn't get freed by jobs that switch to a different
1079 	 * vmspace.
1080 	 */
1081 	p = td->td_proc;
1082 	myvm = vmspace_acquire_ref(p);
1083 
1084 	KASSERT(p->p_textvp == NULL, ("kthread has a textvp"));
1085 
1086 	/*
1087 	 * Allocate and ready the aio control info.  There is one aiop structure
1088 	 * per daemon.
1089 	 */
1090 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
1091 	aiop->aioproc = p;
1092 	aiop->aioprocflags = 0;
1093 
1094 	/*
1095 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1096 	 * and creating too many daemons.)
1097 	 */
1098 	sema_post(&aio_newproc_sem);
1099 
1100 	mtx_lock(&aio_job_mtx);
1101 	for (;;) {
1102 		/*
1103 		 * Take daemon off of free queue
1104 		 */
1105 		if (aiop->aioprocflags & AIOP_FREE) {
1106 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1107 			aiop->aioprocflags &= ~AIOP_FREE;
1108 		}
1109 
1110 		/*
1111 		 * Check for jobs.
1112 		 */
1113 		while ((job = aio_selectjob(aiop)) != NULL) {
1114 			mtx_unlock(&aio_job_mtx);
1115 
1116 			ki = job->userproc->p_aioinfo;
1117 			job->handle_fn(job);
1118 
1119 			mtx_lock(&aio_job_mtx);
1120 			/* Decrement the active job count. */
1121 			ki->kaio_active_count--;
1122 		}
1123 
1124 		/*
1125 		 * Disconnect from user address space.
1126 		 */
1127 		if (p->p_vmspace != myvm) {
1128 			mtx_unlock(&aio_job_mtx);
1129 			vmspace_switch_aio(myvm);
1130 			mtx_lock(&aio_job_mtx);
1131 			/*
1132 			 * We have to restart to avoid race, we only sleep if
1133 			 * no job can be selected.
1134 			 */
1135 			continue;
1136 		}
1137 
1138 		mtx_assert(&aio_job_mtx, MA_OWNED);
1139 
1140 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1141 		aiop->aioprocflags |= AIOP_FREE;
1142 
1143 		/*
1144 		 * If daemon is inactive for a long time, allow it to exit,
1145 		 * thereby freeing resources.
1146 		 */
1147 		if (msleep(p, &aio_job_mtx, PRIBIO, "aiordy",
1148 		    aiod_lifetime) == EWOULDBLOCK && TAILQ_EMPTY(&aio_jobs) &&
1149 		    (aiop->aioprocflags & AIOP_FREE) &&
1150 		    num_aio_procs > target_aio_procs)
1151 			break;
1152 	}
1153 	TAILQ_REMOVE(&aio_freeproc, aiop, list);
1154 	num_aio_procs--;
1155 	mtx_unlock(&aio_job_mtx);
1156 	uma_zfree(aiop_zone, aiop);
1157 	free_unr(aiod_unr, id);
1158 	vmspace_free(myvm);
1159 
1160 	KASSERT(p->p_vmspace == myvm,
1161 	    ("AIOD: bad vmspace for exiting daemon"));
1162 	KASSERT(myvm->vm_refcnt > 1,
1163 	    ("AIOD: bad vm refcnt for exiting daemon: %d", myvm->vm_refcnt));
1164 	kproc_exit(0);
1165 }
1166 
1167 /*
1168  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1169  * AIO daemon modifies its environment itself.
1170  */
1171 static int
1172 aio_newproc(int *start)
1173 {
1174 	int error;
1175 	struct proc *p;
1176 	int id;
1177 
1178 	id = alloc_unr(aiod_unr);
1179 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1180 		RFNOWAIT, 0, "aiod%d", id);
1181 	if (error == 0) {
1182 		/*
1183 		 * Wait until daemon is started.
1184 		 */
1185 		sema_wait(&aio_newproc_sem);
1186 		mtx_lock(&aio_job_mtx);
1187 		num_aio_procs++;
1188 		if (start != NULL)
1189 			(*start)--;
1190 		mtx_unlock(&aio_job_mtx);
1191 	} else {
1192 		free_unr(aiod_unr, id);
1193 	}
1194 	return (error);
1195 }
1196 
1197 /*
1198  * Try the high-performance, low-overhead bio method for eligible
1199  * VCHR devices.  This method doesn't use an aio helper thread, and
1200  * thus has very low overhead.
1201  *
1202  * Assumes that the caller, aio_aqueue(), has incremented the file
1203  * structure's reference count, preventing its deallocation for the
1204  * duration of this call.
1205  */
1206 static int
1207 aio_qbio(struct proc *p, struct kaiocb *job)
1208 {
1209 	struct aiocb *cb;
1210 	struct file *fp;
1211 	struct bio *bp;
1212 	struct buf *pbuf;
1213 	struct vnode *vp;
1214 	struct cdevsw *csw;
1215 	struct cdev *dev;
1216 	struct kaioinfo *ki;
1217 	int error, ref, poff;
1218 	vm_prot_t prot;
1219 
1220 	cb = &job->uaiocb;
1221 	fp = job->fd_file;
1222 
1223 	if (!(cb->aio_lio_opcode == LIO_WRITE ||
1224 	    cb->aio_lio_opcode == LIO_READ))
1225 		return (-1);
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 >= max_buf_aio) {
1266 			error = EAGAIN;
1267 			goto unref;
1268 		}
1269 
1270 		job->pbuf = pbuf = uma_zalloc(pbuf_zone, M_WAITOK);
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_biowakeup;
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 		atomic_add_int(&num_unmapped_aio, 1);
1309 	}
1310 
1311 	/* Perform transfer. */
1312 	csw->d_strategy(bp);
1313 	dev_relthread(dev, ref);
1314 	return (0);
1315 
1316 doerror:
1317 	if (pbuf != NULL) {
1318 		AIO_LOCK(ki);
1319 		ki->kaio_buffer_count--;
1320 		AIO_UNLOCK(ki);
1321 		uma_zfree(pbuf_zone, pbuf);
1322 		job->pbuf = NULL;
1323 	}
1324 	g_destroy_bio(bp);
1325 	job->bp = NULL;
1326 unref:
1327 	dev_relthread(dev, ref);
1328 	return (error);
1329 }
1330 
1331 #ifdef COMPAT_FREEBSD6
1332 static int
1333 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1334 {
1335 
1336 	/*
1337 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1338 	 * supported by AIO with the old sigevent structure.
1339 	 */
1340 	nsig->sigev_notify = osig->sigev_notify;
1341 	switch (nsig->sigev_notify) {
1342 	case SIGEV_NONE:
1343 		break;
1344 	case SIGEV_SIGNAL:
1345 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1346 		break;
1347 	case SIGEV_KEVENT:
1348 		nsig->sigev_notify_kqueue =
1349 		    osig->__sigev_u.__sigev_notify_kqueue;
1350 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1351 		break;
1352 	default:
1353 		return (EINVAL);
1354 	}
1355 	return (0);
1356 }
1357 
1358 static int
1359 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1360 {
1361 	struct oaiocb *ojob;
1362 	int error;
1363 
1364 	bzero(kjob, sizeof(struct aiocb));
1365 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1366 	if (error)
1367 		return (error);
1368 	ojob = (struct oaiocb *)kjob;
1369 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1370 }
1371 #endif
1372 
1373 static int
1374 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1375 {
1376 
1377 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1378 }
1379 
1380 static long
1381 aiocb_fetch_status(struct aiocb *ujob)
1382 {
1383 
1384 	return (fuword(&ujob->_aiocb_private.status));
1385 }
1386 
1387 static long
1388 aiocb_fetch_error(struct aiocb *ujob)
1389 {
1390 
1391 	return (fuword(&ujob->_aiocb_private.error));
1392 }
1393 
1394 static int
1395 aiocb_store_status(struct aiocb *ujob, long status)
1396 {
1397 
1398 	return (suword(&ujob->_aiocb_private.status, status));
1399 }
1400 
1401 static int
1402 aiocb_store_error(struct aiocb *ujob, long error)
1403 {
1404 
1405 	return (suword(&ujob->_aiocb_private.error, error));
1406 }
1407 
1408 static int
1409 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1410 {
1411 
1412 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1413 }
1414 
1415 static int
1416 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1417 {
1418 
1419 	return (suword(ujobp, (long)ujob));
1420 }
1421 
1422 static struct aiocb_ops aiocb_ops = {
1423 	.aio_copyin = aiocb_copyin,
1424 	.fetch_status = aiocb_fetch_status,
1425 	.fetch_error = aiocb_fetch_error,
1426 	.store_status = aiocb_store_status,
1427 	.store_error = aiocb_store_error,
1428 	.store_kernelinfo = aiocb_store_kernelinfo,
1429 	.store_aiocb = aiocb_store_aiocb,
1430 };
1431 
1432 #ifdef COMPAT_FREEBSD6
1433 static struct aiocb_ops aiocb_ops_osigevent = {
1434 	.aio_copyin = aiocb_copyin_old_sigevent,
1435 	.fetch_status = aiocb_fetch_status,
1436 	.fetch_error = aiocb_fetch_error,
1437 	.store_status = aiocb_store_status,
1438 	.store_error = aiocb_store_error,
1439 	.store_kernelinfo = aiocb_store_kernelinfo,
1440 	.store_aiocb = aiocb_store_aiocb,
1441 };
1442 #endif
1443 
1444 /*
1445  * Queue a new AIO request.  Choosing either the threaded or direct bio VCHR
1446  * technique is done in this code.
1447  */
1448 int
1449 aio_aqueue(struct thread *td, struct aiocb *ujob, struct aioliojob *lj,
1450     int type, struct aiocb_ops *ops)
1451 {
1452 	struct proc *p = td->td_proc;
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 >= max_aio_queue_per_proc) {
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->aio_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, &cap_pwrite_rights, &fp);
1531 		break;
1532 	case LIO_READ:
1533 		error = fget_read(td, fd, &cap_pread_rights, &fp);
1534 		break;
1535 	case LIO_SYNC:
1536 		error = fget(td, fd, &cap_fsync_rights, &fp);
1537 		break;
1538 	case LIO_MLOCK:
1539 		fp = NULL;
1540 		break;
1541 	case LIO_NOP:
1542 		error = fget(td, fd, &cap_no_rights, &fp);
1543 		break;
1544 	default:
1545 		error = EINVAL;
1546 	}
1547 	if (error) {
1548 		uma_zfree(aiocb_zone, job);
1549 		ops->store_error(ujob, error);
1550 		return (error);
1551 	}
1552 
1553 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1554 		error = EINVAL;
1555 		goto aqueue_fail;
1556 	}
1557 
1558 	if ((opcode == LIO_READ || opcode == LIO_WRITE) &&
1559 	    job->uaiocb.aio_offset < 0 &&
1560 	    (fp->f_vnode == NULL || fp->f_vnode->v_type != VCHR)) {
1561 		error = EINVAL;
1562 		goto aqueue_fail;
1563 	}
1564 
1565 	job->fd_file = fp;
1566 
1567 	mtx_lock(&aio_job_mtx);
1568 	jid = jobrefid++;
1569 	job->seqno = jobseqno++;
1570 	mtx_unlock(&aio_job_mtx);
1571 	error = ops->store_kernelinfo(ujob, jid);
1572 	if (error) {
1573 		error = EINVAL;
1574 		goto aqueue_fail;
1575 	}
1576 	job->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1577 
1578 	if (opcode == LIO_NOP) {
1579 		fdrop(fp, td);
1580 		uma_zfree(aiocb_zone, job);
1581 		return (0);
1582 	}
1583 
1584 	if (job->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1585 		goto no_kqueue;
1586 	evflags = job->uaiocb.aio_sigevent.sigev_notify_kevent_flags;
1587 	if ((evflags & ~(EV_CLEAR | EV_DISPATCH | EV_ONESHOT)) != 0) {
1588 		error = EINVAL;
1589 		goto aqueue_fail;
1590 	}
1591 	kqfd = job->uaiocb.aio_sigevent.sigev_notify_kqueue;
1592 	memset(&kev, 0, sizeof(kev));
1593 	kev.ident = (uintptr_t)job->ujob;
1594 	kev.filter = EVFILT_AIO;
1595 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1 | evflags;
1596 	kev.data = (intptr_t)job;
1597 	kev.udata = job->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1598 	error = kqfd_register(kqfd, &kev, td, M_WAITOK);
1599 	if (error)
1600 		goto aqueue_fail;
1601 
1602 no_kqueue:
1603 
1604 	ops->store_error(ujob, EINPROGRESS);
1605 	job->uaiocb._aiocb_private.error = EINPROGRESS;
1606 	job->userproc = p;
1607 	job->cred = crhold(td->td_ucred);
1608 	job->jobflags = KAIOCB_QUEUEING;
1609 	job->lio = lj;
1610 
1611 	if (opcode == LIO_MLOCK) {
1612 		aio_schedule(job, aio_process_mlock);
1613 		error = 0;
1614 	} else if (fp->f_ops->fo_aio_queue == NULL)
1615 		error = aio_queue_file(fp, job);
1616 	else
1617 		error = fo_aio_queue(fp, job);
1618 	if (error)
1619 		goto aqueue_fail;
1620 
1621 	AIO_LOCK(ki);
1622 	job->jobflags &= ~KAIOCB_QUEUEING;
1623 	TAILQ_INSERT_TAIL(&ki->kaio_all, job, allist);
1624 	ki->kaio_count++;
1625 	if (lj)
1626 		lj->lioj_count++;
1627 	atomic_add_int(&num_queue_count, 1);
1628 	if (job->jobflags & KAIOCB_FINISHED) {
1629 		/*
1630 		 * The queue callback completed the request synchronously.
1631 		 * The bulk of the completion is deferred in that case
1632 		 * until this point.
1633 		 */
1634 		aio_bio_done_notify(p, job);
1635 	} else
1636 		TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, job, plist);
1637 	AIO_UNLOCK(ki);
1638 	return (0);
1639 
1640 aqueue_fail:
1641 	knlist_delete(&job->klist, curthread, 0);
1642 	if (fp)
1643 		fdrop(fp, td);
1644 	uma_zfree(aiocb_zone, job);
1645 	ops->store_error(ujob, error);
1646 	return (error);
1647 }
1648 
1649 static void
1650 aio_cancel_daemon_job(struct kaiocb *job)
1651 {
1652 
1653 	mtx_lock(&aio_job_mtx);
1654 	if (!aio_cancel_cleared(job))
1655 		TAILQ_REMOVE(&aio_jobs, job, list);
1656 	mtx_unlock(&aio_job_mtx);
1657 	aio_cancel(job);
1658 }
1659 
1660 void
1661 aio_schedule(struct kaiocb *job, aio_handle_fn_t *func)
1662 {
1663 
1664 	mtx_lock(&aio_job_mtx);
1665 	if (!aio_set_cancel_function(job, aio_cancel_daemon_job)) {
1666 		mtx_unlock(&aio_job_mtx);
1667 		aio_cancel(job);
1668 		return;
1669 	}
1670 	job->handle_fn = func;
1671 	TAILQ_INSERT_TAIL(&aio_jobs, job, list);
1672 	aio_kick_nowait(job->userproc);
1673 	mtx_unlock(&aio_job_mtx);
1674 }
1675 
1676 static void
1677 aio_cancel_sync(struct kaiocb *job)
1678 {
1679 	struct kaioinfo *ki;
1680 
1681 	ki = job->userproc->p_aioinfo;
1682 	AIO_LOCK(ki);
1683 	if (!aio_cancel_cleared(job))
1684 		TAILQ_REMOVE(&ki->kaio_syncqueue, job, list);
1685 	AIO_UNLOCK(ki);
1686 	aio_cancel(job);
1687 }
1688 
1689 int
1690 aio_queue_file(struct file *fp, struct kaiocb *job)
1691 {
1692 	struct kaioinfo *ki;
1693 	struct kaiocb *job2;
1694 	struct vnode *vp;
1695 	struct mount *mp;
1696 	int error;
1697 	bool safe;
1698 
1699 	ki = job->userproc->p_aioinfo;
1700 	error = aio_qbio(job->userproc, job);
1701 	if (error >= 0)
1702 		return (error);
1703 	safe = false;
1704 	if (fp->f_type == DTYPE_VNODE) {
1705 		vp = fp->f_vnode;
1706 		if (vp->v_type == VREG || vp->v_type == VDIR) {
1707 			mp = fp->f_vnode->v_mount;
1708 			if (mp == NULL || (mp->mnt_flag & MNT_LOCAL) != 0)
1709 				safe = true;
1710 		}
1711 	}
1712 	if (!(safe || enable_aio_unsafe)) {
1713 		counted_warning(&unsafe_warningcnt,
1714 		    "is attempting to use unsafe AIO requests");
1715 		return (EOPNOTSUPP);
1716 	}
1717 
1718 	switch (job->uaiocb.aio_lio_opcode) {
1719 	case LIO_READ:
1720 	case LIO_WRITE:
1721 		aio_schedule(job, aio_process_rw);
1722 		error = 0;
1723 		break;
1724 	case LIO_SYNC:
1725 		AIO_LOCK(ki);
1726 		TAILQ_FOREACH(job2, &ki->kaio_jobqueue, plist) {
1727 			if (job2->fd_file == job->fd_file &&
1728 			    job2->uaiocb.aio_lio_opcode != LIO_SYNC &&
1729 			    job2->seqno < job->seqno) {
1730 				job2->jobflags |= KAIOCB_CHECKSYNC;
1731 				job->pending++;
1732 			}
1733 		}
1734 		if (job->pending != 0) {
1735 			if (!aio_set_cancel_function_locked(job,
1736 				aio_cancel_sync)) {
1737 				AIO_UNLOCK(ki);
1738 				aio_cancel(job);
1739 				return (0);
1740 			}
1741 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, job, list);
1742 			AIO_UNLOCK(ki);
1743 			return (0);
1744 		}
1745 		AIO_UNLOCK(ki);
1746 		aio_schedule(job, aio_process_sync);
1747 		error = 0;
1748 		break;
1749 	default:
1750 		error = EINVAL;
1751 	}
1752 	return (error);
1753 }
1754 
1755 static void
1756 aio_kick_nowait(struct proc *userp)
1757 {
1758 	struct kaioinfo *ki = userp->p_aioinfo;
1759 	struct aioproc *aiop;
1760 
1761 	mtx_assert(&aio_job_mtx, MA_OWNED);
1762 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1763 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1764 		aiop->aioprocflags &= ~AIOP_FREE;
1765 		wakeup(aiop->aioproc);
1766 	} else if (num_aio_resv_start + num_aio_procs < max_aio_procs &&
1767 	    ki->kaio_active_count + num_aio_resv_start < max_aio_per_proc) {
1768 		taskqueue_enqueue(taskqueue_aiod_kick, &ki->kaio_task);
1769 	}
1770 }
1771 
1772 static int
1773 aio_kick(struct proc *userp)
1774 {
1775 	struct kaioinfo *ki = userp->p_aioinfo;
1776 	struct aioproc *aiop;
1777 	int error, ret = 0;
1778 
1779 	mtx_assert(&aio_job_mtx, MA_OWNED);
1780 retryproc:
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 < max_aio_per_proc) {
1787 		num_aio_resv_start++;
1788 		mtx_unlock(&aio_job_mtx);
1789 		error = aio_newproc(&num_aio_resv_start);
1790 		mtx_lock(&aio_job_mtx);
1791 		if (error) {
1792 			num_aio_resv_start--;
1793 			goto retryproc;
1794 		}
1795 	} else {
1796 		ret = -1;
1797 	}
1798 	return (ret);
1799 }
1800 
1801 static void
1802 aio_kick_helper(void *context, int pending)
1803 {
1804 	struct proc *userp = context;
1805 
1806 	mtx_lock(&aio_job_mtx);
1807 	while (--pending >= 0) {
1808 		if (aio_kick(userp))
1809 			break;
1810 	}
1811 	mtx_unlock(&aio_job_mtx);
1812 }
1813 
1814 /*
1815  * Support the aio_return system call, as a side-effect, kernel resources are
1816  * released.
1817  */
1818 static int
1819 kern_aio_return(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
1820 {
1821 	struct proc *p = td->td_proc;
1822 	struct kaiocb *job;
1823 	struct kaioinfo *ki;
1824 	long status, error;
1825 
1826 	ki = p->p_aioinfo;
1827 	if (ki == NULL)
1828 		return (EINVAL);
1829 	AIO_LOCK(ki);
1830 	TAILQ_FOREACH(job, &ki->kaio_done, plist) {
1831 		if (job->ujob == ujob)
1832 			break;
1833 	}
1834 	if (job != NULL) {
1835 		MPASS(job->jobflags & KAIOCB_FINISHED);
1836 		status = job->uaiocb._aiocb_private.status;
1837 		error = job->uaiocb._aiocb_private.error;
1838 		td->td_retval[0] = status;
1839 		td->td_ru.ru_oublock += job->outblock;
1840 		td->td_ru.ru_inblock += job->inblock;
1841 		td->td_ru.ru_msgsnd += job->msgsnd;
1842 		td->td_ru.ru_msgrcv += job->msgrcv;
1843 		aio_free_entry(job);
1844 		AIO_UNLOCK(ki);
1845 		ops->store_error(ujob, error);
1846 		ops->store_status(ujob, status);
1847 	} else {
1848 		error = EINVAL;
1849 		AIO_UNLOCK(ki);
1850 	}
1851 	return (error);
1852 }
1853 
1854 int
1855 sys_aio_return(struct thread *td, struct aio_return_args *uap)
1856 {
1857 
1858 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1859 }
1860 
1861 /*
1862  * Allow a process to wakeup when any of the I/O requests are completed.
1863  */
1864 static int
1865 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1866     struct timespec *ts)
1867 {
1868 	struct proc *p = td->td_proc;
1869 	struct timeval atv;
1870 	struct kaioinfo *ki;
1871 	struct kaiocb *firstjob, *job;
1872 	int error, i, timo;
1873 
1874 	timo = 0;
1875 	if (ts) {
1876 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1877 			return (EINVAL);
1878 
1879 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1880 		if (itimerfix(&atv))
1881 			return (EINVAL);
1882 		timo = tvtohz(&atv);
1883 	}
1884 
1885 	ki = p->p_aioinfo;
1886 	if (ki == NULL)
1887 		return (EAGAIN);
1888 
1889 	if (njoblist == 0)
1890 		return (0);
1891 
1892 	AIO_LOCK(ki);
1893 	for (;;) {
1894 		firstjob = NULL;
1895 		error = 0;
1896 		TAILQ_FOREACH(job, &ki->kaio_all, allist) {
1897 			for (i = 0; i < njoblist; i++) {
1898 				if (job->ujob == ujoblist[i]) {
1899 					if (firstjob == NULL)
1900 						firstjob = job;
1901 					if (job->jobflags & KAIOCB_FINISHED)
1902 						goto RETURN;
1903 				}
1904 			}
1905 		}
1906 		/* All tasks were finished. */
1907 		if (firstjob == NULL)
1908 			break;
1909 
1910 		ki->kaio_flags |= KAIO_WAKEUP;
1911 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1912 		    "aiospn", timo);
1913 		if (error == ERESTART)
1914 			error = EINTR;
1915 		if (error)
1916 			break;
1917 	}
1918 RETURN:
1919 	AIO_UNLOCK(ki);
1920 	return (error);
1921 }
1922 
1923 int
1924 sys_aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1925 {
1926 	struct timespec ts, *tsp;
1927 	struct aiocb **ujoblist;
1928 	int error;
1929 
1930 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
1931 		return (EINVAL);
1932 
1933 	if (uap->timeout) {
1934 		/* Get timespec struct. */
1935 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1936 			return (error);
1937 		tsp = &ts;
1938 	} else
1939 		tsp = NULL;
1940 
1941 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
1942 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1943 	if (error == 0)
1944 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1945 	free(ujoblist, M_AIOS);
1946 	return (error);
1947 }
1948 
1949 /*
1950  * aio_cancel cancels any non-bio aio operations not currently in progress.
1951  */
1952 int
1953 sys_aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1954 {
1955 	struct proc *p = td->td_proc;
1956 	struct kaioinfo *ki;
1957 	struct kaiocb *job, *jobn;
1958 	struct file *fp;
1959 	int error;
1960 	int cancelled = 0;
1961 	int notcancelled = 0;
1962 	struct vnode *vp;
1963 
1964 	/* Lookup file object. */
1965 	error = fget(td, uap->fd, &cap_no_rights, &fp);
1966 	if (error)
1967 		return (error);
1968 
1969 	ki = p->p_aioinfo;
1970 	if (ki == NULL)
1971 		goto done;
1972 
1973 	if (fp->f_type == DTYPE_VNODE) {
1974 		vp = fp->f_vnode;
1975 		if (vn_isdisk(vp, &error)) {
1976 			fdrop(fp, td);
1977 			td->td_retval[0] = AIO_NOTCANCELED;
1978 			return (0);
1979 		}
1980 	}
1981 
1982 	AIO_LOCK(ki);
1983 	TAILQ_FOREACH_SAFE(job, &ki->kaio_jobqueue, plist, jobn) {
1984 		if ((uap->fd == job->uaiocb.aio_fildes) &&
1985 		    ((uap->aiocbp == NULL) ||
1986 		     (uap->aiocbp == job->ujob))) {
1987 			if (aio_cancel_job(p, ki, job)) {
1988 				cancelled++;
1989 			} else {
1990 				notcancelled++;
1991 			}
1992 			if (uap->aiocbp != NULL)
1993 				break;
1994 		}
1995 	}
1996 	AIO_UNLOCK(ki);
1997 
1998 done:
1999 	fdrop(fp, td);
2000 
2001 	if (uap->aiocbp != NULL) {
2002 		if (cancelled) {
2003 			td->td_retval[0] = AIO_CANCELED;
2004 			return (0);
2005 		}
2006 	}
2007 
2008 	if (notcancelled) {
2009 		td->td_retval[0] = AIO_NOTCANCELED;
2010 		return (0);
2011 	}
2012 
2013 	if (cancelled) {
2014 		td->td_retval[0] = AIO_CANCELED;
2015 		return (0);
2016 	}
2017 
2018 	td->td_retval[0] = AIO_ALLDONE;
2019 
2020 	return (0);
2021 }
2022 
2023 /*
2024  * aio_error is implemented in the kernel level for compatibility purposes
2025  * only.  For a user mode async implementation, it would be best to do it in
2026  * a userland subroutine.
2027  */
2028 static int
2029 kern_aio_error(struct thread *td, struct aiocb *ujob, struct aiocb_ops *ops)
2030 {
2031 	struct proc *p = td->td_proc;
2032 	struct kaiocb *job;
2033 	struct kaioinfo *ki;
2034 	int status;
2035 
2036 	ki = p->p_aioinfo;
2037 	if (ki == NULL) {
2038 		td->td_retval[0] = EINVAL;
2039 		return (0);
2040 	}
2041 
2042 	AIO_LOCK(ki);
2043 	TAILQ_FOREACH(job, &ki->kaio_all, allist) {
2044 		if (job->ujob == ujob) {
2045 			if (job->jobflags & KAIOCB_FINISHED)
2046 				td->td_retval[0] =
2047 					job->uaiocb._aiocb_private.error;
2048 			else
2049 				td->td_retval[0] = EINPROGRESS;
2050 			AIO_UNLOCK(ki);
2051 			return (0);
2052 		}
2053 	}
2054 	AIO_UNLOCK(ki);
2055 
2056 	/*
2057 	 * Hack for failure of aio_aqueue.
2058 	 */
2059 	status = ops->fetch_status(ujob);
2060 	if (status == -1) {
2061 		td->td_retval[0] = ops->fetch_error(ujob);
2062 		return (0);
2063 	}
2064 
2065 	td->td_retval[0] = EINVAL;
2066 	return (0);
2067 }
2068 
2069 int
2070 sys_aio_error(struct thread *td, struct aio_error_args *uap)
2071 {
2072 
2073 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2074 }
2075 
2076 /* syscall - asynchronous read from a file (REALTIME) */
2077 #ifdef COMPAT_FREEBSD6
2078 int
2079 freebsd6_aio_read(struct thread *td, struct freebsd6_aio_read_args *uap)
2080 {
2081 
2082 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2083 	    &aiocb_ops_osigevent));
2084 }
2085 #endif
2086 
2087 int
2088 sys_aio_read(struct thread *td, struct aio_read_args *uap)
2089 {
2090 
2091 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2092 }
2093 
2094 /* syscall - asynchronous write to a file (REALTIME) */
2095 #ifdef COMPAT_FREEBSD6
2096 int
2097 freebsd6_aio_write(struct thread *td, struct freebsd6_aio_write_args *uap)
2098 {
2099 
2100 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2101 	    &aiocb_ops_osigevent));
2102 }
2103 #endif
2104 
2105 int
2106 sys_aio_write(struct thread *td, struct aio_write_args *uap)
2107 {
2108 
2109 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2110 }
2111 
2112 int
2113 sys_aio_mlock(struct thread *td, struct aio_mlock_args *uap)
2114 {
2115 
2116 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_MLOCK, &aiocb_ops));
2117 }
2118 
2119 static int
2120 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2121     struct aiocb **acb_list, int nent, struct sigevent *sig,
2122     struct aiocb_ops *ops)
2123 {
2124 	struct proc *p = td->td_proc;
2125 	struct aiocb *job;
2126 	struct kaioinfo *ki;
2127 	struct aioliojob *lj;
2128 	struct kevent kev;
2129 	int error;
2130 	int nagain, nerror;
2131 	int i;
2132 
2133 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2134 		return (EINVAL);
2135 
2136 	if (nent < 0 || nent > max_aio_queue_per_proc)
2137 		return (EINVAL);
2138 
2139 	if (p->p_aioinfo == NULL)
2140 		aio_init_aioinfo(p);
2141 
2142 	ki = p->p_aioinfo;
2143 
2144 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2145 	lj->lioj_flags = 0;
2146 	lj->lioj_count = 0;
2147 	lj->lioj_finished_count = 0;
2148 	knlist_init_mtx(&lj->klist, AIO_MTX(ki));
2149 	ksiginfo_init(&lj->lioj_ksi);
2150 
2151 	/*
2152 	 * Setup signal.
2153 	 */
2154 	if (sig && (mode == LIO_NOWAIT)) {
2155 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2156 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2157 			/* Assume only new style KEVENT */
2158 			memset(&kev, 0, sizeof(kev));
2159 			kev.filter = EVFILT_LIO;
2160 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2161 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2162 			kev.data = (intptr_t)lj;
2163 			/* pass user defined sigval data */
2164 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2165 			error = kqfd_register(
2166 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td,
2167 			    M_WAITOK);
2168 			if (error) {
2169 				uma_zfree(aiolio_zone, lj);
2170 				return (error);
2171 			}
2172 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2173 			;
2174 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2175 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2176 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2177 					uma_zfree(aiolio_zone, lj);
2178 					return EINVAL;
2179 				}
2180 				lj->lioj_flags |= LIOJ_SIGNAL;
2181 		} else {
2182 			uma_zfree(aiolio_zone, lj);
2183 			return EINVAL;
2184 		}
2185 	}
2186 
2187 	AIO_LOCK(ki);
2188 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2189 	/*
2190 	 * Add extra aiocb count to avoid the lio to be freed
2191 	 * by other threads doing aio_waitcomplete or aio_return,
2192 	 * and prevent event from being sent until we have queued
2193 	 * all tasks.
2194 	 */
2195 	lj->lioj_count = 1;
2196 	AIO_UNLOCK(ki);
2197 
2198 	/*
2199 	 * Get pointers to the list of I/O requests.
2200 	 */
2201 	nagain = 0;
2202 	nerror = 0;
2203 	for (i = 0; i < nent; i++) {
2204 		job = acb_list[i];
2205 		if (job != NULL) {
2206 			error = aio_aqueue(td, job, lj, LIO_NOP, ops);
2207 			if (error == EAGAIN)
2208 				nagain++;
2209 			else if (error != 0)
2210 				nerror++;
2211 		}
2212 	}
2213 
2214 	error = 0;
2215 	AIO_LOCK(ki);
2216 	if (mode == LIO_WAIT) {
2217 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2218 			ki->kaio_flags |= KAIO_WAKEUP;
2219 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2220 			    PRIBIO | PCATCH, "aiospn", 0);
2221 			if (error == ERESTART)
2222 				error = EINTR;
2223 			if (error)
2224 				break;
2225 		}
2226 	} else {
2227 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2228 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2229 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2230 				KNOTE_LOCKED(&lj->klist, 1);
2231 			}
2232 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2233 			    == LIOJ_SIGNAL
2234 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2235 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2236 				aio_sendsig(p, &lj->lioj_signal,
2237 					    &lj->lioj_ksi);
2238 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2239 			}
2240 		}
2241 	}
2242 	lj->lioj_count--;
2243 	if (lj->lioj_count == 0) {
2244 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2245 		knlist_delete(&lj->klist, curthread, 1);
2246 		PROC_LOCK(p);
2247 		sigqueue_take(&lj->lioj_ksi);
2248 		PROC_UNLOCK(p);
2249 		AIO_UNLOCK(ki);
2250 		uma_zfree(aiolio_zone, lj);
2251 	} else
2252 		AIO_UNLOCK(ki);
2253 
2254 	if (nerror)
2255 		return (EIO);
2256 	else if (nagain)
2257 		return (EAGAIN);
2258 	else
2259 		return (error);
2260 }
2261 
2262 /* syscall - list directed I/O (REALTIME) */
2263 #ifdef COMPAT_FREEBSD6
2264 int
2265 freebsd6_lio_listio(struct thread *td, struct freebsd6_lio_listio_args *uap)
2266 {
2267 	struct aiocb **acb_list;
2268 	struct sigevent *sigp, sig;
2269 	struct osigevent osig;
2270 	int error, nent;
2271 
2272 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2273 		return (EINVAL);
2274 
2275 	nent = uap->nent;
2276 	if (nent < 0 || nent > max_aio_queue_per_proc)
2277 		return (EINVAL);
2278 
2279 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2280 		error = copyin(uap->sig, &osig, sizeof(osig));
2281 		if (error)
2282 			return (error);
2283 		error = convert_old_sigevent(&osig, &sig);
2284 		if (error)
2285 			return (error);
2286 		sigp = &sig;
2287 	} else
2288 		sigp = NULL;
2289 
2290 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2291 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2292 	if (error == 0)
2293 		error = kern_lio_listio(td, uap->mode,
2294 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2295 		    &aiocb_ops_osigevent);
2296 	free(acb_list, M_LIO);
2297 	return (error);
2298 }
2299 #endif
2300 
2301 /* syscall - list directed I/O (REALTIME) */
2302 int
2303 sys_lio_listio(struct thread *td, struct lio_listio_args *uap)
2304 {
2305 	struct aiocb **acb_list;
2306 	struct sigevent *sigp, sig;
2307 	int error, nent;
2308 
2309 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2310 		return (EINVAL);
2311 
2312 	nent = uap->nent;
2313 	if (nent < 0 || nent > max_aio_queue_per_proc)
2314 		return (EINVAL);
2315 
2316 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2317 		error = copyin(uap->sig, &sig, sizeof(sig));
2318 		if (error)
2319 			return (error);
2320 		sigp = &sig;
2321 	} else
2322 		sigp = NULL;
2323 
2324 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2325 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2326 	if (error == 0)
2327 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2328 		    nent, sigp, &aiocb_ops);
2329 	free(acb_list, M_LIO);
2330 	return (error);
2331 }
2332 
2333 static void
2334 aio_biowakeup(struct bio *bp)
2335 {
2336 	struct kaiocb *job = (struct kaiocb *)bp->bio_caller1;
2337 	struct proc *userp;
2338 	struct kaioinfo *ki;
2339 	size_t nbytes;
2340 	int error, nblks;
2341 
2342 	/* Release mapping into kernel space. */
2343 	userp = job->userproc;
2344 	ki = userp->p_aioinfo;
2345 	if (job->pbuf) {
2346 		pmap_qremove((vm_offset_t)job->pbuf->b_data, job->npages);
2347 		uma_zfree(pbuf_zone, job->pbuf);
2348 		job->pbuf = NULL;
2349 		atomic_subtract_int(&num_buf_aio, 1);
2350 		AIO_LOCK(ki);
2351 		ki->kaio_buffer_count--;
2352 		AIO_UNLOCK(ki);
2353 	} else
2354 		atomic_subtract_int(&num_unmapped_aio, 1);
2355 	vm_page_unhold_pages(job->pages, job->npages);
2356 
2357 	bp = job->bp;
2358 	job->bp = NULL;
2359 	nbytes = job->uaiocb.aio_nbytes - bp->bio_resid;
2360 	error = 0;
2361 	if (bp->bio_flags & BIO_ERROR)
2362 		error = bp->bio_error;
2363 	nblks = btodb(nbytes);
2364 	if (job->uaiocb.aio_lio_opcode == LIO_WRITE)
2365 		job->outblock += nblks;
2366 	else
2367 		job->inblock += nblks;
2368 
2369 	if (error)
2370 		aio_complete(job, -1, error);
2371 	else
2372 		aio_complete(job, nbytes, 0);
2373 
2374 	g_destroy_bio(bp);
2375 }
2376 
2377 /* syscall - wait for the next completion of an aio request */
2378 static int
2379 kern_aio_waitcomplete(struct thread *td, struct aiocb **ujobp,
2380     struct timespec *ts, struct aiocb_ops *ops)
2381 {
2382 	struct proc *p = td->td_proc;
2383 	struct timeval atv;
2384 	struct kaioinfo *ki;
2385 	struct kaiocb *job;
2386 	struct aiocb *ujob;
2387 	long error, status;
2388 	int timo;
2389 
2390 	ops->store_aiocb(ujobp, NULL);
2391 
2392 	if (ts == NULL) {
2393 		timo = 0;
2394 	} else if (ts->tv_sec == 0 && ts->tv_nsec == 0) {
2395 		timo = -1;
2396 	} else {
2397 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2398 			return (EINVAL);
2399 
2400 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2401 		if (itimerfix(&atv))
2402 			return (EINVAL);
2403 		timo = tvtohz(&atv);
2404 	}
2405 
2406 	if (p->p_aioinfo == NULL)
2407 		aio_init_aioinfo(p);
2408 	ki = p->p_aioinfo;
2409 
2410 	error = 0;
2411 	job = NULL;
2412 	AIO_LOCK(ki);
2413 	while ((job = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2414 		if (timo == -1) {
2415 			error = EWOULDBLOCK;
2416 			break;
2417 		}
2418 		ki->kaio_flags |= KAIO_WAKEUP;
2419 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2420 		    "aiowc", timo);
2421 		if (timo && error == ERESTART)
2422 			error = EINTR;
2423 		if (error)
2424 			break;
2425 	}
2426 
2427 	if (job != NULL) {
2428 		MPASS(job->jobflags & KAIOCB_FINISHED);
2429 		ujob = job->ujob;
2430 		status = job->uaiocb._aiocb_private.status;
2431 		error = job->uaiocb._aiocb_private.error;
2432 		td->td_retval[0] = status;
2433 		td->td_ru.ru_oublock += job->outblock;
2434 		td->td_ru.ru_inblock += job->inblock;
2435 		td->td_ru.ru_msgsnd += job->msgsnd;
2436 		td->td_ru.ru_msgrcv += job->msgrcv;
2437 		aio_free_entry(job);
2438 		AIO_UNLOCK(ki);
2439 		ops->store_aiocb(ujobp, ujob);
2440 		ops->store_error(ujob, error);
2441 		ops->store_status(ujob, status);
2442 	} else
2443 		AIO_UNLOCK(ki);
2444 
2445 	return (error);
2446 }
2447 
2448 int
2449 sys_aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2450 {
2451 	struct timespec ts, *tsp;
2452 	int error;
2453 
2454 	if (uap->timeout) {
2455 		/* Get timespec struct. */
2456 		error = copyin(uap->timeout, &ts, sizeof(ts));
2457 		if (error)
2458 			return (error);
2459 		tsp = &ts;
2460 	} else
2461 		tsp = NULL;
2462 
2463 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2464 }
2465 
2466 static int
2467 kern_aio_fsync(struct thread *td, int op, struct aiocb *ujob,
2468     struct aiocb_ops *ops)
2469 {
2470 
2471 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2472 		return (EINVAL);
2473 	return (aio_aqueue(td, ujob, NULL, LIO_SYNC, ops));
2474 }
2475 
2476 int
2477 sys_aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2478 {
2479 
2480 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2481 }
2482 
2483 /* kqueue attach function */
2484 static int
2485 filt_aioattach(struct knote *kn)
2486 {
2487 	struct kaiocb *job;
2488 
2489 	job = (struct kaiocb *)(uintptr_t)kn->kn_sdata;
2490 
2491 	/*
2492 	 * The job pointer must be validated before using it, so
2493 	 * registration is restricted to the kernel; the user cannot
2494 	 * set EV_FLAG1.
2495 	 */
2496 	if ((kn->kn_flags & EV_FLAG1) == 0)
2497 		return (EPERM);
2498 	kn->kn_ptr.p_aio = job;
2499 	kn->kn_flags &= ~EV_FLAG1;
2500 
2501 	knlist_add(&job->klist, kn, 0);
2502 
2503 	return (0);
2504 }
2505 
2506 /* kqueue detach function */
2507 static void
2508 filt_aiodetach(struct knote *kn)
2509 {
2510 	struct knlist *knl;
2511 
2512 	knl = &kn->kn_ptr.p_aio->klist;
2513 	knl->kl_lock(knl->kl_lockarg);
2514 	if (!knlist_empty(knl))
2515 		knlist_remove(knl, kn, 1);
2516 	knl->kl_unlock(knl->kl_lockarg);
2517 }
2518 
2519 /* kqueue filter function */
2520 /*ARGSUSED*/
2521 static int
2522 filt_aio(struct knote *kn, long hint)
2523 {
2524 	struct kaiocb *job = kn->kn_ptr.p_aio;
2525 
2526 	kn->kn_data = job->uaiocb._aiocb_private.error;
2527 	if (!(job->jobflags & KAIOCB_FINISHED))
2528 		return (0);
2529 	kn->kn_flags |= EV_EOF;
2530 	return (1);
2531 }
2532 
2533 /* kqueue attach function */
2534 static int
2535 filt_lioattach(struct knote *kn)
2536 {
2537 	struct aioliojob *lj;
2538 
2539 	lj = (struct aioliojob *)(uintptr_t)kn->kn_sdata;
2540 
2541 	/*
2542 	 * The aioliojob pointer must be validated before using it, so
2543 	 * registration is restricted to the kernel; the user cannot
2544 	 * set EV_FLAG1.
2545 	 */
2546 	if ((kn->kn_flags & EV_FLAG1) == 0)
2547 		return (EPERM);
2548 	kn->kn_ptr.p_lio = lj;
2549 	kn->kn_flags &= ~EV_FLAG1;
2550 
2551 	knlist_add(&lj->klist, kn, 0);
2552 
2553 	return (0);
2554 }
2555 
2556 /* kqueue detach function */
2557 static void
2558 filt_liodetach(struct knote *kn)
2559 {
2560 	struct knlist *knl;
2561 
2562 	knl = &kn->kn_ptr.p_lio->klist;
2563 	knl->kl_lock(knl->kl_lockarg);
2564 	if (!knlist_empty(knl))
2565 		knlist_remove(knl, kn, 1);
2566 	knl->kl_unlock(knl->kl_lockarg);
2567 }
2568 
2569 /* kqueue filter function */
2570 /*ARGSUSED*/
2571 static int
2572 filt_lio(struct knote *kn, long hint)
2573 {
2574 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2575 
2576 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2577 }
2578 
2579 #ifdef COMPAT_FREEBSD32
2580 #include <sys/mount.h>
2581 #include <sys/socket.h>
2582 #include <compat/freebsd32/freebsd32.h>
2583 #include <compat/freebsd32/freebsd32_proto.h>
2584 #include <compat/freebsd32/freebsd32_signal.h>
2585 #include <compat/freebsd32/freebsd32_syscall.h>
2586 #include <compat/freebsd32/freebsd32_util.h>
2587 
2588 struct __aiocb_private32 {
2589 	int32_t	status;
2590 	int32_t	error;
2591 	uint32_t kernelinfo;
2592 };
2593 
2594 #ifdef COMPAT_FREEBSD6
2595 typedef struct oaiocb32 {
2596 	int	aio_fildes;		/* File descriptor */
2597 	uint64_t aio_offset __packed;	/* File offset for I/O */
2598 	uint32_t aio_buf;		/* I/O buffer in process space */
2599 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2600 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2601 	int	aio_lio_opcode;		/* LIO opcode */
2602 	int	aio_reqprio;		/* Request priority -- ignored */
2603 	struct	__aiocb_private32 _aiocb_private;
2604 } oaiocb32_t;
2605 #endif
2606 
2607 typedef struct aiocb32 {
2608 	int32_t	aio_fildes;		/* File descriptor */
2609 	uint64_t aio_offset __packed;	/* File offset for I/O */
2610 	uint32_t aio_buf;		/* I/O buffer in process space */
2611 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2612 	int	__spare__[2];
2613 	uint32_t __spare2__;
2614 	int	aio_lio_opcode;		/* LIO opcode */
2615 	int	aio_reqprio;		/* Request priority -- ignored */
2616 	struct	__aiocb_private32 _aiocb_private;
2617 	struct	sigevent32 aio_sigevent;	/* Signal to deliver */
2618 } aiocb32_t;
2619 
2620 #ifdef COMPAT_FREEBSD6
2621 static int
2622 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2623 {
2624 
2625 	/*
2626 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2627 	 * supported by AIO with the old sigevent structure.
2628 	 */
2629 	CP(*osig, *nsig, sigev_notify);
2630 	switch (nsig->sigev_notify) {
2631 	case SIGEV_NONE:
2632 		break;
2633 	case SIGEV_SIGNAL:
2634 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2635 		break;
2636 	case SIGEV_KEVENT:
2637 		nsig->sigev_notify_kqueue =
2638 		    osig->__sigev_u.__sigev_notify_kqueue;
2639 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2640 		break;
2641 	default:
2642 		return (EINVAL);
2643 	}
2644 	return (0);
2645 }
2646 
2647 static int
2648 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2649 {
2650 	struct oaiocb32 job32;
2651 	int error;
2652 
2653 	bzero(kjob, sizeof(struct aiocb));
2654 	error = copyin(ujob, &job32, sizeof(job32));
2655 	if (error)
2656 		return (error);
2657 
2658 	CP(job32, *kjob, aio_fildes);
2659 	CP(job32, *kjob, aio_offset);
2660 	PTRIN_CP(job32, *kjob, aio_buf);
2661 	CP(job32, *kjob, aio_nbytes);
2662 	CP(job32, *kjob, aio_lio_opcode);
2663 	CP(job32, *kjob, aio_reqprio);
2664 	CP(job32, *kjob, _aiocb_private.status);
2665 	CP(job32, *kjob, _aiocb_private.error);
2666 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2667 	return (convert_old_sigevent32(&job32.aio_sigevent,
2668 	    &kjob->aio_sigevent));
2669 }
2670 #endif
2671 
2672 static int
2673 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2674 {
2675 	struct aiocb32 job32;
2676 	int error;
2677 
2678 	error = copyin(ujob, &job32, sizeof(job32));
2679 	if (error)
2680 		return (error);
2681 	CP(job32, *kjob, aio_fildes);
2682 	CP(job32, *kjob, aio_offset);
2683 	PTRIN_CP(job32, *kjob, aio_buf);
2684 	CP(job32, *kjob, aio_nbytes);
2685 	CP(job32, *kjob, aio_lio_opcode);
2686 	CP(job32, *kjob, aio_reqprio);
2687 	CP(job32, *kjob, _aiocb_private.status);
2688 	CP(job32, *kjob, _aiocb_private.error);
2689 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2690 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2691 }
2692 
2693 static long
2694 aiocb32_fetch_status(struct aiocb *ujob)
2695 {
2696 	struct aiocb32 *ujob32;
2697 
2698 	ujob32 = (struct aiocb32 *)ujob;
2699 	return (fuword32(&ujob32->_aiocb_private.status));
2700 }
2701 
2702 static long
2703 aiocb32_fetch_error(struct aiocb *ujob)
2704 {
2705 	struct aiocb32 *ujob32;
2706 
2707 	ujob32 = (struct aiocb32 *)ujob;
2708 	return (fuword32(&ujob32->_aiocb_private.error));
2709 }
2710 
2711 static int
2712 aiocb32_store_status(struct aiocb *ujob, long status)
2713 {
2714 	struct aiocb32 *ujob32;
2715 
2716 	ujob32 = (struct aiocb32 *)ujob;
2717 	return (suword32(&ujob32->_aiocb_private.status, status));
2718 }
2719 
2720 static int
2721 aiocb32_store_error(struct aiocb *ujob, long error)
2722 {
2723 	struct aiocb32 *ujob32;
2724 
2725 	ujob32 = (struct aiocb32 *)ujob;
2726 	return (suword32(&ujob32->_aiocb_private.error, error));
2727 }
2728 
2729 static int
2730 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2731 {
2732 	struct aiocb32 *ujob32;
2733 
2734 	ujob32 = (struct aiocb32 *)ujob;
2735 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2736 }
2737 
2738 static int
2739 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2740 {
2741 
2742 	return (suword32(ujobp, (long)ujob));
2743 }
2744 
2745 static struct aiocb_ops aiocb32_ops = {
2746 	.aio_copyin = aiocb32_copyin,
2747 	.fetch_status = aiocb32_fetch_status,
2748 	.fetch_error = aiocb32_fetch_error,
2749 	.store_status = aiocb32_store_status,
2750 	.store_error = aiocb32_store_error,
2751 	.store_kernelinfo = aiocb32_store_kernelinfo,
2752 	.store_aiocb = aiocb32_store_aiocb,
2753 };
2754 
2755 #ifdef COMPAT_FREEBSD6
2756 static struct aiocb_ops aiocb32_ops_osigevent = {
2757 	.aio_copyin = aiocb32_copyin_old_sigevent,
2758 	.fetch_status = aiocb32_fetch_status,
2759 	.fetch_error = aiocb32_fetch_error,
2760 	.store_status = aiocb32_store_status,
2761 	.store_error = aiocb32_store_error,
2762 	.store_kernelinfo = aiocb32_store_kernelinfo,
2763 	.store_aiocb = aiocb32_store_aiocb,
2764 };
2765 #endif
2766 
2767 int
2768 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2769 {
2770 
2771 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2772 }
2773 
2774 int
2775 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2776 {
2777 	struct timespec32 ts32;
2778 	struct timespec ts, *tsp;
2779 	struct aiocb **ujoblist;
2780 	uint32_t *ujoblist32;
2781 	int error, i;
2782 
2783 	if (uap->nent < 0 || uap->nent > max_aio_queue_per_proc)
2784 		return (EINVAL);
2785 
2786 	if (uap->timeout) {
2787 		/* Get timespec struct. */
2788 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2789 			return (error);
2790 		CP(ts32, ts, tv_sec);
2791 		CP(ts32, ts, tv_nsec);
2792 		tsp = &ts;
2793 	} else
2794 		tsp = NULL;
2795 
2796 	ujoblist = malloc(uap->nent * sizeof(ujoblist[0]), M_AIOS, M_WAITOK);
2797 	ujoblist32 = (uint32_t *)ujoblist;
2798 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2799 	    sizeof(ujoblist32[0]));
2800 	if (error == 0) {
2801 		for (i = uap->nent - 1; i >= 0; i--)
2802 			ujoblist[i] = PTRIN(ujoblist32[i]);
2803 
2804 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2805 	}
2806 	free(ujoblist, M_AIOS);
2807 	return (error);
2808 }
2809 
2810 int
2811 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2812 {
2813 
2814 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2815 }
2816 
2817 #ifdef COMPAT_FREEBSD6
2818 int
2819 freebsd6_freebsd32_aio_read(struct thread *td,
2820     struct freebsd6_freebsd32_aio_read_args *uap)
2821 {
2822 
2823 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2824 	    &aiocb32_ops_osigevent));
2825 }
2826 #endif
2827 
2828 int
2829 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2830 {
2831 
2832 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2833 	    &aiocb32_ops));
2834 }
2835 
2836 #ifdef COMPAT_FREEBSD6
2837 int
2838 freebsd6_freebsd32_aio_write(struct thread *td,
2839     struct freebsd6_freebsd32_aio_write_args *uap)
2840 {
2841 
2842 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2843 	    &aiocb32_ops_osigevent));
2844 }
2845 #endif
2846 
2847 int
2848 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2849 {
2850 
2851 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2852 	    &aiocb32_ops));
2853 }
2854 
2855 int
2856 freebsd32_aio_mlock(struct thread *td, struct freebsd32_aio_mlock_args *uap)
2857 {
2858 
2859 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_MLOCK,
2860 	    &aiocb32_ops));
2861 }
2862 
2863 int
2864 freebsd32_aio_waitcomplete(struct thread *td,
2865     struct freebsd32_aio_waitcomplete_args *uap)
2866 {
2867 	struct timespec32 ts32;
2868 	struct timespec ts, *tsp;
2869 	int error;
2870 
2871 	if (uap->timeout) {
2872 		/* Get timespec struct. */
2873 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2874 		if (error)
2875 			return (error);
2876 		CP(ts32, ts, tv_sec);
2877 		CP(ts32, ts, tv_nsec);
2878 		tsp = &ts;
2879 	} else
2880 		tsp = NULL;
2881 
2882 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2883 	    &aiocb32_ops));
2884 }
2885 
2886 int
2887 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2888 {
2889 
2890 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2891 	    &aiocb32_ops));
2892 }
2893 
2894 #ifdef COMPAT_FREEBSD6
2895 int
2896 freebsd6_freebsd32_lio_listio(struct thread *td,
2897     struct freebsd6_freebsd32_lio_listio_args *uap)
2898 {
2899 	struct aiocb **acb_list;
2900 	struct sigevent *sigp, sig;
2901 	struct osigevent32 osig;
2902 	uint32_t *acb_list32;
2903 	int error, i, nent;
2904 
2905 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2906 		return (EINVAL);
2907 
2908 	nent = uap->nent;
2909 	if (nent < 0 || nent > max_aio_queue_per_proc)
2910 		return (EINVAL);
2911 
2912 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2913 		error = copyin(uap->sig, &osig, sizeof(osig));
2914 		if (error)
2915 			return (error);
2916 		error = convert_old_sigevent32(&osig, &sig);
2917 		if (error)
2918 			return (error);
2919 		sigp = &sig;
2920 	} else
2921 		sigp = NULL;
2922 
2923 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2924 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2925 	if (error) {
2926 		free(acb_list32, M_LIO);
2927 		return (error);
2928 	}
2929 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2930 	for (i = 0; i < nent; i++)
2931 		acb_list[i] = PTRIN(acb_list32[i]);
2932 	free(acb_list32, M_LIO);
2933 
2934 	error = kern_lio_listio(td, uap->mode,
2935 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2936 	    &aiocb32_ops_osigevent);
2937 	free(acb_list, M_LIO);
2938 	return (error);
2939 }
2940 #endif
2941 
2942 int
2943 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2944 {
2945 	struct aiocb **acb_list;
2946 	struct sigevent *sigp, sig;
2947 	struct sigevent32 sig32;
2948 	uint32_t *acb_list32;
2949 	int error, i, nent;
2950 
2951 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2952 		return (EINVAL);
2953 
2954 	nent = uap->nent;
2955 	if (nent < 0 || nent > max_aio_queue_per_proc)
2956 		return (EINVAL);
2957 
2958 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2959 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2960 		if (error)
2961 			return (error);
2962 		error = convert_sigevent32(&sig32, &sig);
2963 		if (error)
2964 			return (error);
2965 		sigp = &sig;
2966 	} else
2967 		sigp = NULL;
2968 
2969 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2970 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2971 	if (error) {
2972 		free(acb_list32, M_LIO);
2973 		return (error);
2974 	}
2975 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2976 	for (i = 0; i < nent; i++)
2977 		acb_list[i] = PTRIN(acb_list32[i]);
2978 	free(acb_list32, M_LIO);
2979 
2980 	error = kern_lio_listio(td, uap->mode,
2981 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2982 	    &aiocb32_ops);
2983 	free(acb_list, M_LIO);
2984 	return (error);
2985 }
2986 
2987 #endif
2988