xref: /freebsd/sys/kern/vfs_aio.c (revision b3aaa0cc21c63d388230c7ef2a80abd631ff20d5)
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/eventhandler.h>
32 #include <sys/sysproto.h>
33 #include <sys/filedesc.h>
34 #include <sys/kernel.h>
35 #include <sys/module.h>
36 #include <sys/kthread.h>
37 #include <sys/fcntl.h>
38 #include <sys/file.h>
39 #include <sys/limits.h>
40 #include <sys/lock.h>
41 #include <sys/mutex.h>
42 #include <sys/unistd.h>
43 #include <sys/posix4.h>
44 #include <sys/proc.h>
45 #include <sys/resourcevar.h>
46 #include <sys/signalvar.h>
47 #include <sys/protosw.h>
48 #include <sys/sema.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/syscall.h>
52 #include <sys/sysent.h>
53 #include <sys/sysctl.h>
54 #include <sys/sx.h>
55 #include <sys/taskqueue.h>
56 #include <sys/vnode.h>
57 #include <sys/conf.h>
58 #include <sys/event.h>
59 #include <sys/mount.h>
60 
61 #include <machine/atomic.h>
62 
63 #include <vm/vm.h>
64 #include <vm/vm_extern.h>
65 #include <vm/pmap.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_object.h>
68 #include <vm/uma.h>
69 #include <sys/aio.h>
70 
71 #include "opt_vfs_aio.h"
72 
73 /*
74  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
75  * overflow. (XXX will be removed soon.)
76  */
77 static u_long jobrefid;
78 
79 /*
80  * Counter for aio_fsync.
81  */
82 static uint64_t jobseqno;
83 
84 #define JOBST_NULL		0
85 #define JOBST_JOBQSOCK		1
86 #define JOBST_JOBQGLOBAL	2
87 #define JOBST_JOBRUNNING	3
88 #define JOBST_JOBFINISHED	4
89 #define JOBST_JOBQBUF		5
90 #define JOBST_JOBQSYNC		6
91 
92 #ifndef MAX_AIO_PER_PROC
93 #define MAX_AIO_PER_PROC	32
94 #endif
95 
96 #ifndef MAX_AIO_QUEUE_PER_PROC
97 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
98 #endif
99 
100 #ifndef MAX_AIO_PROCS
101 #define MAX_AIO_PROCS		32
102 #endif
103 
104 #ifndef MAX_AIO_QUEUE
105 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
106 #endif
107 
108 #ifndef TARGET_AIO_PROCS
109 #define TARGET_AIO_PROCS	4
110 #endif
111 
112 #ifndef MAX_BUF_AIO
113 #define MAX_BUF_AIO		16
114 #endif
115 
116 #ifndef AIOD_TIMEOUT_DEFAULT
117 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
118 #endif
119 
120 #ifndef AIOD_LIFETIME_DEFAULT
121 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
122 #endif
123 
124 FEATURE(aio, "Asynchronous I/O");
125 
126 static MALLOC_DEFINE(M_LIO, "lio", "listio aio control block list");
127 
128 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
129 
130 static int max_aio_procs = MAX_AIO_PROCS;
131 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
132 	CTLFLAG_RW, &max_aio_procs, 0,
133 	"Maximum number of kernel threads to use for handling async IO ");
134 
135 static int num_aio_procs = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
137 	CTLFLAG_RD, &num_aio_procs, 0,
138 	"Number of presently active kernel threads for async IO");
139 
140 /*
141  * The code will adjust the actual number of AIO processes towards this
142  * number when it gets a chance.
143  */
144 static int target_aio_procs = TARGET_AIO_PROCS;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
146 	0, "Preferred number of ready kernel threads for async IO");
147 
148 static int max_queue_count = MAX_AIO_QUEUE;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
150     "Maximum number of aio requests to queue, globally");
151 
152 static int num_queue_count = 0;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
154     "Number of queued aio requests");
155 
156 static int num_buf_aio = 0;
157 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
158     "Number of aio requests presently handled by the buf subsystem");
159 
160 /* Number of async I/O thread in the process of being started */
161 /* XXX This should be local to aio_aqueue() */
162 static int num_aio_resv_start = 0;
163 
164 static int aiod_timeout;
165 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
166     "Timeout value for synchronous aio operations");
167 
168 static int aiod_lifetime;
169 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
170     "Maximum lifetime for idle aiod");
171 
172 static int unloadable = 0;
173 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
174     "Allow unload of aio (not recommended)");
175 
176 
177 static int max_aio_per_proc = MAX_AIO_PER_PROC;
178 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
179     0, "Maximum active aio requests per process (stored in the process)");
180 
181 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
182 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
183     &max_aio_queue_per_proc, 0,
184     "Maximum queued aio requests per process (stored in the process)");
185 
186 static int max_buf_aio = MAX_BUF_AIO;
187 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
188     "Maximum buf aio requests per process (stored in the process)");
189 
190 typedef struct oaiocb {
191 	int	aio_fildes;		/* File descriptor */
192 	off_t	aio_offset;		/* File offset for I/O */
193 	volatile void *aio_buf;         /* I/O buffer in process space */
194 	size_t	aio_nbytes;		/* Number of bytes for I/O */
195 	struct	osigevent aio_sigevent;	/* Signal to deliver */
196 	int	aio_lio_opcode;		/* LIO opcode */
197 	int	aio_reqprio;		/* Request priority -- ignored */
198 	struct	__aiocb_private	_aiocb_private;
199 } oaiocb_t;
200 
201 /*
202  * Below is a key of locks used to protect each member of struct aiocblist
203  * aioliojob and kaioinfo and any backends.
204  *
205  * * - need not protected
206  * a - locked by kaioinfo lock
207  * b - locked by backend lock, the backend lock can be null in some cases,
208  *     for example, BIO belongs to this type, in this case, proc lock is
209  *     reused.
210  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
211  */
212 
213 /*
214  * Current, there is only two backends: BIO and generic file I/O.
215  * socket I/O is served by generic file I/O, this is not a good idea, since
216  * disk file I/O and any other types without O_NONBLOCK flag can block daemon
217  * threads, if there is no thread to serve socket I/O, the socket I/O will be
218  * delayed too long or starved, we should create some threads dedicated to
219  * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
220  * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
221  * structure is not safe because there is race between userland and aio
222  * daemons.
223  */
224 
225 struct aiocblist {
226 	TAILQ_ENTRY(aiocblist) list;	/* (b) internal list of for backend */
227 	TAILQ_ENTRY(aiocblist) plist;	/* (a) list of jobs for each backend */
228 	TAILQ_ENTRY(aiocblist) allist;  /* (a) list of all jobs in proc */
229 	int	jobflags;		/* (a) job flags */
230 	int	jobstate;		/* (b) job state */
231 	int	inputcharge;		/* (*) input blockes */
232 	int	outputcharge;		/* (*) output blockes */
233 	struct	buf *bp;		/* (*) private to BIO backend,
234 				  	 * buffer pointer
235 					 */
236 	struct	proc *userproc;		/* (*) user process */
237 	struct  ucred *cred;		/* (*) active credential when created */
238 	struct	file *fd_file;		/* (*) pointer to file structure */
239 	struct	aioliojob *lio;		/* (*) optional lio job */
240 	struct	aiocb *uuaiocb;		/* (*) pointer in userspace of aiocb */
241 	struct	knlist klist;		/* (a) list of knotes */
242 	struct	aiocb uaiocb;		/* (*) kernel I/O control block */
243 	ksiginfo_t ksi;			/* (a) realtime signal info */
244 	struct	task biotask;		/* (*) private to BIO backend */
245 	uint64_t seqno;			/* (*) job number */
246 	int	pending;		/* (a) number of pending I/O, aio_fsync only */
247 };
248 
249 /* jobflags */
250 #define AIOCBLIST_DONE		0x01
251 #define AIOCBLIST_BUFDONE	0x02
252 #define AIOCBLIST_RUNDOWN	0x04
253 #define AIOCBLIST_CHECKSYNC	0x08
254 
255 /*
256  * AIO process info
257  */
258 #define AIOP_FREE	0x1			/* proc on free queue */
259 
260 struct aiothreadlist {
261 	int aiothreadflags;			/* (c) AIO proc flags */
262 	TAILQ_ENTRY(aiothreadlist) list;	/* (c) list of processes */
263 	struct thread *aiothread;		/* (*) the AIO thread */
264 };
265 
266 /*
267  * data-structure for lio signal management
268  */
269 struct aioliojob {
270 	int	lioj_flags;			/* (a) listio flags */
271 	int	lioj_count;			/* (a) listio flags */
272 	int	lioj_finished_count;		/* (a) listio flags */
273 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
274 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
275 	struct  knlist klist;			/* (a) list of knotes */
276 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
277 };
278 
279 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
280 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
281 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
282 
283 /*
284  * per process aio data structure
285  */
286 struct kaioinfo {
287 	struct mtx	kaio_mtx;	/* the lock to protect this struct */
288 	int	kaio_flags;		/* (a) per process kaio flags */
289 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
290 	int	kaio_active_count;	/* (c) number of currently used AIOs */
291 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
292 	int	kaio_count;		/* (a) size of AIO queue */
293 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
294 	int	kaio_buffer_count;	/* (a) number of physio buffers */
295 	TAILQ_HEAD(,aiocblist) kaio_all;	/* (a) all AIOs in the process */
296 	TAILQ_HEAD(,aiocblist) kaio_done;	/* (a) done queue for process */
297 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
298 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* (a) job queue for process */
299 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* (a) buffer job queue for process */
300 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;  /* (a) queue for aios waiting on sockets,
301 						 *  NOT USED YET.
302 						 */
303 	TAILQ_HEAD(,aiocblist) kaio_syncqueue;	/* (a) queue for aio_fsync */
304 	struct	task	kaio_task;	/* (*) task to kick aio threads */
305 };
306 
307 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
308 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
309 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
310 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
311 
312 #define KAIO_RUNDOWN	0x1	/* process is being run down */
313 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
314 
315 /*
316  * Operations used to interact with userland aio control blocks.
317  * Different ABIs provide their own operations.
318  */
319 struct aiocb_ops {
320 	int	(*copyin)(struct aiocb *ujob, struct aiocb *kjob);
321 	long	(*fetch_status)(struct aiocb *ujob);
322 	long	(*fetch_error)(struct aiocb *ujob);
323 	int	(*store_status)(struct aiocb *ujob, long status);
324 	int	(*store_error)(struct aiocb *ujob, long error);
325 	int	(*store_kernelinfo)(struct aiocb *ujob, long jobref);
326 	int	(*store_aiocb)(struct aiocb **ujobp, struct aiocb *ujob);
327 };
328 
329 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* (c) Idle daemons */
330 static struct sema aio_newproc_sem;
331 static struct mtx aio_job_mtx;
332 static struct mtx aio_sock_mtx;
333 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* (c) Async job list */
334 static struct unrhdr *aiod_unr;
335 
336 void		aio_init_aioinfo(struct proc *p);
337 static void	aio_onceonly(void);
338 static int	aio_free_entry(struct aiocblist *aiocbe);
339 static void	aio_process(struct aiocblist *aiocbe);
340 static int	aio_newproc(int *);
341 int		aio_aqueue(struct thread *td, struct aiocb *job,
342 			struct aioliojob *lio, int type, struct aiocb_ops *ops);
343 static void	aio_physwakeup(struct buf *bp);
344 static void	aio_proc_rundown(void *arg, struct proc *p);
345 static void	aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
346 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
347 static void	biohelper(void *, int);
348 static void	aio_daemon(void *param);
349 static void	aio_swake_cb(struct socket *, struct sockbuf *);
350 static int	aio_unload(void);
351 static void	aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
352 #define DONE_BUF	1
353 #define DONE_QUEUE	2
354 static int	aio_kick(struct proc *userp);
355 static void	aio_kick_nowait(struct proc *userp);
356 static void	aio_kick_helper(void *context, int pending);
357 static int	filt_aioattach(struct knote *kn);
358 static void	filt_aiodetach(struct knote *kn);
359 static int	filt_aio(struct knote *kn, long hint);
360 static int	filt_lioattach(struct knote *kn);
361 static void	filt_liodetach(struct knote *kn);
362 static int	filt_lio(struct knote *kn, long hint);
363 
364 /*
365  * Zones for:
366  * 	kaio	Per process async io info
367  *	aiop	async io thread data
368  *	aiocb	async io jobs
369  *	aiol	list io job pointer - internal to aio_suspend XXX
370  *	aiolio	list io jobs
371  */
372 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
373 
374 /* kqueue filters for aio */
375 static struct filterops aio_filtops =
376 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
377 static struct filterops lio_filtops =
378 	{ 0, filt_lioattach, filt_liodetach, filt_lio };
379 
380 static eventhandler_tag exit_tag, exec_tag;
381 
382 TASKQUEUE_DEFINE_THREAD(aiod_bio);
383 
384 /*
385  * Main operations function for use as a kernel module.
386  */
387 static int
388 aio_modload(struct module *module, int cmd, void *arg)
389 {
390 	int error = 0;
391 
392 	switch (cmd) {
393 	case MOD_LOAD:
394 		aio_onceonly();
395 		break;
396 	case MOD_UNLOAD:
397 		error = aio_unload();
398 		break;
399 	case MOD_SHUTDOWN:
400 		break;
401 	default:
402 		error = EINVAL;
403 		break;
404 	}
405 	return (error);
406 }
407 
408 static moduledata_t aio_mod = {
409 	"aio",
410 	&aio_modload,
411 	NULL
412 };
413 
414 SYSCALL_MODULE_HELPER(aio_cancel);
415 SYSCALL_MODULE_HELPER(aio_error);
416 SYSCALL_MODULE_HELPER(aio_fsync);
417 SYSCALL_MODULE_HELPER(aio_read);
418 SYSCALL_MODULE_HELPER(aio_return);
419 SYSCALL_MODULE_HELPER(aio_suspend);
420 SYSCALL_MODULE_HELPER(aio_waitcomplete);
421 SYSCALL_MODULE_HELPER(aio_write);
422 SYSCALL_MODULE_HELPER(lio_listio);
423 SYSCALL_MODULE_HELPER(oaio_read);
424 SYSCALL_MODULE_HELPER(oaio_write);
425 SYSCALL_MODULE_HELPER(olio_listio);
426 
427 DECLARE_MODULE(aio, aio_mod,
428 	SI_SUB_VFS, SI_ORDER_ANY);
429 MODULE_VERSION(aio, 1);
430 
431 /*
432  * Startup initialization
433  */
434 static void
435 aio_onceonly(void)
436 {
437 
438 	/* XXX: should probably just use so->callback */
439 	aio_swake = &aio_swake_cb;
440 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
441 	    EVENTHANDLER_PRI_ANY);
442 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
443 	    EVENTHANDLER_PRI_ANY);
444 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
445 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
446 	TAILQ_INIT(&aio_freeproc);
447 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
448 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
449 	mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
450 	TAILQ_INIT(&aio_jobs);
451 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
452 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
453 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
454 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
455 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
456 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
457 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
458 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
459 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
460 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
461 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
462 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
463 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
464 	jobrefid = 1;
465 	async_io_version = _POSIX_VERSION;
466 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
467 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
468 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
469 }
470 
471 /*
472  * Callback for unload of AIO when used as a module.
473  */
474 static int
475 aio_unload(void)
476 {
477 	int error;
478 
479 	/*
480 	 * XXX: no unloads by default, it's too dangerous.
481 	 * perhaps we could do it if locked out callers and then
482 	 * did an aio_proc_rundown() on each process.
483 	 *
484 	 * jhb: aio_proc_rundown() needs to run on curproc though,
485 	 * so I don't think that would fly.
486 	 */
487 	if (!unloadable)
488 		return (EOPNOTSUPP);
489 
490 	error = kqueue_del_filteropts(EVFILT_AIO);
491 	if (error)
492 		return error;
493 	error = kqueue_del_filteropts(EVFILT_LIO);
494 	if (error)
495 		return error;
496 	async_io_version = 0;
497 	aio_swake = NULL;
498 	taskqueue_free(taskqueue_aiod_bio);
499 	delete_unrhdr(aiod_unr);
500 	uma_zdestroy(kaio_zone);
501 	uma_zdestroy(aiop_zone);
502 	uma_zdestroy(aiocb_zone);
503 	uma_zdestroy(aiol_zone);
504 	uma_zdestroy(aiolio_zone);
505 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
506 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
507 	mtx_destroy(&aio_job_mtx);
508 	mtx_destroy(&aio_sock_mtx);
509 	sema_destroy(&aio_newproc_sem);
510 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
511 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
512 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
513 	return (0);
514 }
515 
516 /*
517  * Init the per-process aioinfo structure.  The aioinfo limits are set
518  * per-process for user limit (resource) management.
519  */
520 void
521 aio_init_aioinfo(struct proc *p)
522 {
523 	struct kaioinfo *ki;
524 
525 	ki = uma_zalloc(kaio_zone, M_WAITOK);
526 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
527 	ki->kaio_flags = 0;
528 	ki->kaio_maxactive_count = max_aio_per_proc;
529 	ki->kaio_active_count = 0;
530 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
531 	ki->kaio_count = 0;
532 	ki->kaio_ballowed_count = max_buf_aio;
533 	ki->kaio_buffer_count = 0;
534 	TAILQ_INIT(&ki->kaio_all);
535 	TAILQ_INIT(&ki->kaio_done);
536 	TAILQ_INIT(&ki->kaio_jobqueue);
537 	TAILQ_INIT(&ki->kaio_bufqueue);
538 	TAILQ_INIT(&ki->kaio_liojoblist);
539 	TAILQ_INIT(&ki->kaio_sockqueue);
540 	TAILQ_INIT(&ki->kaio_syncqueue);
541 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
542 	PROC_LOCK(p);
543 	if (p->p_aioinfo == NULL) {
544 		p->p_aioinfo = ki;
545 		PROC_UNLOCK(p);
546 	} else {
547 		PROC_UNLOCK(p);
548 		mtx_destroy(&ki->kaio_mtx);
549 		uma_zfree(kaio_zone, ki);
550 	}
551 
552 	while (num_aio_procs < MIN(target_aio_procs, max_aio_procs))
553 		aio_newproc(NULL);
554 }
555 
556 static int
557 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
558 {
559 	int ret = 0;
560 
561 	PROC_LOCK(p);
562 	if (!KSI_ONQ(ksi)) {
563 		ksi->ksi_code = SI_ASYNCIO;
564 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
565 		ret = psignal_event(p, sigev, ksi);
566 	}
567 	PROC_UNLOCK(p);
568 	return (ret);
569 }
570 
571 /*
572  * Free a job entry.  Wait for completion if it is currently active, but don't
573  * delay forever.  If we delay, we return a flag that says that we have to
574  * restart the queue scan.
575  */
576 static int
577 aio_free_entry(struct aiocblist *aiocbe)
578 {
579 	struct kaioinfo *ki;
580 	struct aioliojob *lj;
581 	struct proc *p;
582 
583 	p = aiocbe->userproc;
584 	MPASS(curproc == p);
585 	ki = p->p_aioinfo;
586 	MPASS(ki != NULL);
587 
588 	AIO_LOCK_ASSERT(ki, MA_OWNED);
589 	MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
590 
591 	atomic_subtract_int(&num_queue_count, 1);
592 
593 	ki->kaio_count--;
594 	MPASS(ki->kaio_count >= 0);
595 
596 	TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
597 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
598 
599 	lj = aiocbe->lio;
600 	if (lj) {
601 		lj->lioj_count--;
602 		lj->lioj_finished_count--;
603 
604 		if (lj->lioj_count == 0) {
605 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
606 			/* lio is going away, we need to destroy any knotes */
607 			knlist_delete(&lj->klist, curthread, 1);
608 			PROC_LOCK(p);
609 			sigqueue_take(&lj->lioj_ksi);
610 			PROC_UNLOCK(p);
611 			uma_zfree(aiolio_zone, lj);
612 		}
613 	}
614 
615 	/* aiocbe is going away, we need to destroy any knotes */
616 	knlist_delete(&aiocbe->klist, curthread, 1);
617 	PROC_LOCK(p);
618 	sigqueue_take(&aiocbe->ksi);
619 	PROC_UNLOCK(p);
620 
621 	MPASS(aiocbe->bp == NULL);
622 	aiocbe->jobstate = JOBST_NULL;
623 	AIO_UNLOCK(ki);
624 
625 	/*
626 	 * The thread argument here is used to find the owning process
627 	 * and is also passed to fo_close() which may pass it to various
628 	 * places such as devsw close() routines.  Because of that, we
629 	 * need a thread pointer from the process owning the job that is
630 	 * persistent and won't disappear out from under us or move to
631 	 * another process.
632 	 *
633 	 * Currently, all the callers of this function call it to remove
634 	 * an aiocblist from the current process' job list either via a
635 	 * syscall or due to the current process calling exit() or
636 	 * execve().  Thus, we know that p == curproc.  We also know that
637 	 * curthread can't exit since we are curthread.
638 	 *
639 	 * Therefore, we use curthread as the thread to pass to
640 	 * knlist_delete().  This does mean that it is possible for the
641 	 * thread pointer at close time to differ from the thread pointer
642 	 * at open time, but this is already true of file descriptors in
643 	 * a multithreaded process.
644 	 */
645 	fdrop(aiocbe->fd_file, curthread);
646 	crfree(aiocbe->cred);
647 	uma_zfree(aiocb_zone, aiocbe);
648 	AIO_LOCK(ki);
649 
650 	return (0);
651 }
652 
653 static void
654 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
655 {
656    	aio_proc_rundown(arg, p);
657 }
658 
659 /*
660  * Rundown the jobs for a given process.
661  */
662 static void
663 aio_proc_rundown(void *arg, struct proc *p)
664 {
665 	struct kaioinfo *ki;
666 	struct aioliojob *lj;
667 	struct aiocblist *cbe, *cbn;
668 	struct file *fp;
669 	struct socket *so;
670 	int remove;
671 
672 	KASSERT(curthread->td_proc == p,
673 	    ("%s: called on non-curproc", __func__));
674 	ki = p->p_aioinfo;
675 	if (ki == NULL)
676 		return;
677 
678 	AIO_LOCK(ki);
679 	ki->kaio_flags |= KAIO_RUNDOWN;
680 
681 restart:
682 
683 	/*
684 	 * Try to cancel all pending requests. This code simulates
685 	 * aio_cancel on all pending I/O requests.
686 	 */
687 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
688 		remove = 0;
689 		mtx_lock(&aio_job_mtx);
690 		if (cbe->jobstate == JOBST_JOBQGLOBAL) {
691 			TAILQ_REMOVE(&aio_jobs, cbe, list);
692 			remove = 1;
693 		} else if (cbe->jobstate == JOBST_JOBQSOCK) {
694 			fp = cbe->fd_file;
695 			MPASS(fp->f_type == DTYPE_SOCKET);
696 			so = fp->f_data;
697 			TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
698 			remove = 1;
699 		} else if (cbe->jobstate == JOBST_JOBQSYNC) {
700 			TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
701 			remove = 1;
702 		}
703 		mtx_unlock(&aio_job_mtx);
704 
705 		if (remove) {
706 			cbe->jobstate = JOBST_JOBFINISHED;
707 			cbe->uaiocb._aiocb_private.status = -1;
708 			cbe->uaiocb._aiocb_private.error = ECANCELED;
709 			TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
710 			aio_bio_done_notify(p, cbe, DONE_QUEUE);
711 		}
712 	}
713 
714 	/* Wait for all running I/O to be finished */
715 	if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
716 	    TAILQ_FIRST(&ki->kaio_jobqueue)) {
717 		ki->kaio_flags |= KAIO_WAKEUP;
718 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
719 		goto restart;
720 	}
721 
722 	/* Free all completed I/O requests. */
723 	while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
724 		aio_free_entry(cbe);
725 
726 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
727 		if (lj->lioj_count == 0) {
728 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
729 			knlist_delete(&lj->klist, curthread, 1);
730 			PROC_LOCK(p);
731 			sigqueue_take(&lj->lioj_ksi);
732 			PROC_UNLOCK(p);
733 			uma_zfree(aiolio_zone, lj);
734 		} else {
735 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
736 			    lj->lioj_count, lj->lioj_finished_count);
737 		}
738 	}
739 	AIO_UNLOCK(ki);
740 	taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
741 	mtx_destroy(&ki->kaio_mtx);
742 	uma_zfree(kaio_zone, ki);
743 	p->p_aioinfo = NULL;
744 }
745 
746 /*
747  * Select a job to run (called by an AIO daemon).
748  */
749 static struct aiocblist *
750 aio_selectjob(struct aiothreadlist *aiop)
751 {
752 	struct aiocblist *aiocbe;
753 	struct kaioinfo *ki;
754 	struct proc *userp;
755 
756 	mtx_assert(&aio_job_mtx, MA_OWNED);
757 	TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
758 		userp = aiocbe->userproc;
759 		ki = userp->p_aioinfo;
760 
761 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
762 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
763 			/* Account for currently active jobs. */
764 			ki->kaio_active_count++;
765 			aiocbe->jobstate = JOBST_JOBRUNNING;
766 			break;
767 		}
768 	}
769 	return (aiocbe);
770 }
771 
772 /*
773  *  Move all data to a permanent storage device, this code
774  *  simulates fsync syscall.
775  */
776 static int
777 aio_fsync_vnode(struct thread *td, struct vnode *vp)
778 {
779 	struct mount *mp;
780 	int vfslocked;
781 	int error;
782 
783 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
784 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
785 		goto drop;
786 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
787 	if (vp->v_object != NULL) {
788 		VM_OBJECT_LOCK(vp->v_object);
789 		vm_object_page_clean(vp->v_object, 0, 0, 0);
790 		VM_OBJECT_UNLOCK(vp->v_object);
791 	}
792 	error = VOP_FSYNC(vp, MNT_WAIT, td);
793 
794 	VOP_UNLOCK(vp, 0);
795 	vn_finished_write(mp);
796 drop:
797 	VFS_UNLOCK_GIANT(vfslocked);
798 	return (error);
799 }
800 
801 /*
802  * The AIO processing activity.  This is the code that does the I/O request for
803  * the non-physio version of the operations.  The normal vn operations are used,
804  * and this code should work in all instances for every type of file, including
805  * pipes, sockets, fifos, and regular files.
806  *
807  * XXX I don't think it works well for socket, pipe, and fifo.
808  */
809 static void
810 aio_process(struct aiocblist *aiocbe)
811 {
812 	struct ucred *td_savedcred;
813 	struct thread *td;
814 	struct aiocb *cb;
815 	struct file *fp;
816 	struct socket *so;
817 	struct uio auio;
818 	struct iovec aiov;
819 	int cnt;
820 	int error;
821 	int oublock_st, oublock_end;
822 	int inblock_st, inblock_end;
823 
824 	td = curthread;
825 	td_savedcred = td->td_ucred;
826 	td->td_ucred = aiocbe->cred;
827 	cb = &aiocbe->uaiocb;
828 	fp = aiocbe->fd_file;
829 
830 	if (cb->aio_lio_opcode == LIO_SYNC) {
831 		error = 0;
832 		cnt = 0;
833 		if (fp->f_vnode != NULL)
834 			error = aio_fsync_vnode(td, fp->f_vnode);
835 		cb->_aiocb_private.error = error;
836 		cb->_aiocb_private.status = 0;
837 		td->td_ucred = td_savedcred;
838 		return;
839 	}
840 
841 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
842 	aiov.iov_len = cb->aio_nbytes;
843 
844 	auio.uio_iov = &aiov;
845 	auio.uio_iovcnt = 1;
846 	auio.uio_offset = cb->aio_offset;
847 	auio.uio_resid = cb->aio_nbytes;
848 	cnt = cb->aio_nbytes;
849 	auio.uio_segflg = UIO_USERSPACE;
850 	auio.uio_td = td;
851 
852 	inblock_st = td->td_ru.ru_inblock;
853 	oublock_st = td->td_ru.ru_oublock;
854 	/*
855 	 * aio_aqueue() acquires a reference to the file that is
856 	 * released in aio_free_entry().
857 	 */
858 	if (cb->aio_lio_opcode == LIO_READ) {
859 		auio.uio_rw = UIO_READ;
860 		if (auio.uio_resid == 0)
861 			error = 0;
862 		else
863 			error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
864 	} else {
865 		if (fp->f_type == DTYPE_VNODE)
866 			bwillwrite();
867 		auio.uio_rw = UIO_WRITE;
868 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
869 	}
870 	inblock_end = td->td_ru.ru_inblock;
871 	oublock_end = td->td_ru.ru_oublock;
872 
873 	aiocbe->inputcharge = inblock_end - inblock_st;
874 	aiocbe->outputcharge = oublock_end - oublock_st;
875 
876 	if ((error) && (auio.uio_resid != cnt)) {
877 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
878 			error = 0;
879 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
880 			int sigpipe = 1;
881 			if (fp->f_type == DTYPE_SOCKET) {
882 				so = fp->f_data;
883 				if (so->so_options & SO_NOSIGPIPE)
884 					sigpipe = 0;
885 			}
886 			if (sigpipe) {
887 				PROC_LOCK(aiocbe->userproc);
888 				psignal(aiocbe->userproc, SIGPIPE);
889 				PROC_UNLOCK(aiocbe->userproc);
890 			}
891 		}
892 	}
893 
894 	cnt -= auio.uio_resid;
895 	cb->_aiocb_private.error = error;
896 	cb->_aiocb_private.status = cnt;
897 	td->td_ucred = td_savedcred;
898 }
899 
900 static void
901 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
902 {
903 	struct aioliojob *lj;
904 	struct kaioinfo *ki;
905 	struct aiocblist *scb, *scbn;
906 	int lj_done;
907 
908 	ki = userp->p_aioinfo;
909 	AIO_LOCK_ASSERT(ki, MA_OWNED);
910 	lj = aiocbe->lio;
911 	lj_done = 0;
912 	if (lj) {
913 		lj->lioj_finished_count++;
914 		if (lj->lioj_count == lj->lioj_finished_count)
915 			lj_done = 1;
916 	}
917 	if (type == DONE_QUEUE) {
918 		aiocbe->jobflags |= AIOCBLIST_DONE;
919 	} else {
920 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
921 	}
922 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
923 	aiocbe->jobstate = JOBST_JOBFINISHED;
924 
925 	if (ki->kaio_flags & KAIO_RUNDOWN)
926 		goto notification_done;
927 
928 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
929 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
930 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
931 
932 	KNOTE_LOCKED(&aiocbe->klist, 1);
933 
934 	if (lj_done) {
935 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
936 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
937 			KNOTE_LOCKED(&lj->klist, 1);
938 		}
939 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
940 		    == LIOJ_SIGNAL
941 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
942 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
943 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
944 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
945 		}
946 	}
947 
948 notification_done:
949 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
950 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
951 			if (aiocbe->fd_file == scb->fd_file &&
952 			    aiocbe->seqno < scb->seqno) {
953 				if (--scb->pending == 0) {
954 					mtx_lock(&aio_job_mtx);
955 					scb->jobstate = JOBST_JOBQGLOBAL;
956 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
957 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
958 					aio_kick_nowait(userp);
959 					mtx_unlock(&aio_job_mtx);
960 				}
961 			}
962 		}
963 	}
964 	if (ki->kaio_flags & KAIO_WAKEUP) {
965 		ki->kaio_flags &= ~KAIO_WAKEUP;
966 		wakeup(&userp->p_aioinfo);
967 	}
968 }
969 
970 /*
971  * The AIO daemon, most of the actual work is done in aio_process,
972  * but the setup (and address space mgmt) is done in this routine.
973  */
974 static void
975 aio_daemon(void *_id)
976 {
977 	struct aiocblist *aiocbe;
978 	struct aiothreadlist *aiop;
979 	struct kaioinfo *ki;
980 	struct proc *curcp, *mycp, *userp;
981 	struct vmspace *myvm, *tmpvm;
982 	struct thread *td = curthread;
983 	int id = (intptr_t)_id;
984 
985 	/*
986 	 * Local copies of curproc (cp) and vmspace (myvm)
987 	 */
988 	mycp = td->td_proc;
989 	myvm = mycp->p_vmspace;
990 
991 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
992 
993 	/*
994 	 * Allocate and ready the aio control info.  There is one aiop structure
995 	 * per daemon.
996 	 */
997 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
998 	aiop->aiothread = td;
999 	aiop->aiothreadflags = 0;
1000 
1001 	/* The daemon resides in its own pgrp. */
1002 	setsid(td, NULL);
1003 
1004 	/*
1005 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
1006 	 * and creating too many daemons.)
1007 	 */
1008 	sema_post(&aio_newproc_sem);
1009 
1010 	mtx_lock(&aio_job_mtx);
1011 	for (;;) {
1012 		/*
1013 		 * curcp is the current daemon process context.
1014 		 * userp is the current user process context.
1015 		 */
1016 		curcp = mycp;
1017 
1018 		/*
1019 		 * Take daemon off of free queue
1020 		 */
1021 		if (aiop->aiothreadflags & AIOP_FREE) {
1022 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1023 			aiop->aiothreadflags &= ~AIOP_FREE;
1024 		}
1025 
1026 		/*
1027 		 * Check for jobs.
1028 		 */
1029 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1030 			mtx_unlock(&aio_job_mtx);
1031 			userp = aiocbe->userproc;
1032 
1033 			/*
1034 			 * Connect to process address space for user program.
1035 			 */
1036 			if (userp != curcp) {
1037 				/*
1038 				 * Save the current address space that we are
1039 				 * connected to.
1040 				 */
1041 				tmpvm = mycp->p_vmspace;
1042 
1043 				/*
1044 				 * Point to the new user address space, and
1045 				 * refer to it.
1046 				 */
1047 				mycp->p_vmspace = userp->p_vmspace;
1048 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1049 
1050 				/* Activate the new mapping. */
1051 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1052 
1053 				/*
1054 				 * If the old address space wasn't the daemons
1055 				 * own address space, then we need to remove the
1056 				 * daemon's reference from the other process
1057 				 * that it was acting on behalf of.
1058 				 */
1059 				if (tmpvm != myvm) {
1060 					vmspace_free(tmpvm);
1061 				}
1062 				curcp = userp;
1063 			}
1064 
1065 			ki = userp->p_aioinfo;
1066 
1067 			/* Do the I/O function. */
1068 			aio_process(aiocbe);
1069 
1070 			mtx_lock(&aio_job_mtx);
1071 			/* Decrement the active job count. */
1072 			ki->kaio_active_count--;
1073 			mtx_unlock(&aio_job_mtx);
1074 
1075 			AIO_LOCK(ki);
1076 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1077 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1078 			AIO_UNLOCK(ki);
1079 
1080 			mtx_lock(&aio_job_mtx);
1081 		}
1082 
1083 		/*
1084 		 * Disconnect from user address space.
1085 		 */
1086 		if (curcp != mycp) {
1087 
1088 			mtx_unlock(&aio_job_mtx);
1089 
1090 			/* Get the user address space to disconnect from. */
1091 			tmpvm = mycp->p_vmspace;
1092 
1093 			/* Get original address space for daemon. */
1094 			mycp->p_vmspace = myvm;
1095 
1096 			/* Activate the daemon's address space. */
1097 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1098 #ifdef DIAGNOSTIC
1099 			if (tmpvm == myvm) {
1100 				printf("AIOD: vmspace problem -- %d\n",
1101 				    mycp->p_pid);
1102 			}
1103 #endif
1104 			/* Remove our vmspace reference. */
1105 			vmspace_free(tmpvm);
1106 
1107 			curcp = mycp;
1108 
1109 			mtx_lock(&aio_job_mtx);
1110 			/*
1111 			 * We have to restart to avoid race, we only sleep if
1112 			 * no job can be selected, that should be
1113 			 * curcp == mycp.
1114 			 */
1115 			continue;
1116 		}
1117 
1118 		mtx_assert(&aio_job_mtx, MA_OWNED);
1119 
1120 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1121 		aiop->aiothreadflags |= AIOP_FREE;
1122 
1123 		/*
1124 		 * If daemon is inactive for a long time, allow it to exit,
1125 		 * thereby freeing resources.
1126 		 */
1127 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1128 		    aiod_lifetime)) {
1129 			if (TAILQ_EMPTY(&aio_jobs)) {
1130 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1131 				    (num_aio_procs > target_aio_procs)) {
1132 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1133 					num_aio_procs--;
1134 					mtx_unlock(&aio_job_mtx);
1135 					uma_zfree(aiop_zone, aiop);
1136 					free_unr(aiod_unr, id);
1137 #ifdef DIAGNOSTIC
1138 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1139 						printf("AIOD: bad vm refcnt for"
1140 						    " exiting daemon: %d\n",
1141 						    mycp->p_vmspace->vm_refcnt);
1142 					}
1143 #endif
1144 					kproc_exit(0);
1145 				}
1146 			}
1147 		}
1148 	}
1149 	mtx_unlock(&aio_job_mtx);
1150 	panic("shouldn't be here\n");
1151 }
1152 
1153 /*
1154  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1155  * AIO daemon modifies its environment itself.
1156  */
1157 static int
1158 aio_newproc(int *start)
1159 {
1160 	int error;
1161 	struct proc *p;
1162 	int id;
1163 
1164 	id = alloc_unr(aiod_unr);
1165 	error = kproc_create(aio_daemon, (void *)(intptr_t)id, &p,
1166 		RFNOWAIT, 0, "aiod%d", id);
1167 	if (error == 0) {
1168 		/*
1169 		 * Wait until daemon is started.
1170 		 */
1171 		sema_wait(&aio_newproc_sem);
1172 		mtx_lock(&aio_job_mtx);
1173 		num_aio_procs++;
1174 		if (start != NULL)
1175 			(*start)--;
1176 		mtx_unlock(&aio_job_mtx);
1177 	} else {
1178 		free_unr(aiod_unr, id);
1179 	}
1180 	return (error);
1181 }
1182 
1183 /*
1184  * Try the high-performance, low-overhead physio method for eligible
1185  * VCHR devices.  This method doesn't use an aio helper thread, and
1186  * thus has very low overhead.
1187  *
1188  * Assumes that the caller, aio_aqueue(), has incremented the file
1189  * structure's reference count, preventing its deallocation for the
1190  * duration of this call.
1191  */
1192 static int
1193 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1194 {
1195 	struct aiocb *cb;
1196 	struct file *fp;
1197 	struct buf *bp;
1198 	struct vnode *vp;
1199 	struct kaioinfo *ki;
1200 	struct aioliojob *lj;
1201 	int error;
1202 
1203 	cb = &aiocbe->uaiocb;
1204 	fp = aiocbe->fd_file;
1205 
1206 	if (fp->f_type != DTYPE_VNODE)
1207 		return (-1);
1208 
1209 	vp = fp->f_vnode;
1210 
1211 	/*
1212 	 * If its not a disk, we don't want to return a positive error.
1213 	 * It causes the aio code to not fall through to try the thread
1214 	 * way when you're talking to a regular file.
1215 	 */
1216 	if (!vn_isdisk(vp, &error)) {
1217 		if (error == ENOTBLK)
1218 			return (-1);
1219 		else
1220 			return (error);
1221 	}
1222 
1223 	if (vp->v_bufobj.bo_bsize == 0)
1224 		return (-1);
1225 
1226  	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1227 		return (-1);
1228 
1229 	if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1230 		return (-1);
1231 
1232 	if (cb->aio_nbytes >
1233 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1234 		return (-1);
1235 
1236 	ki = p->p_aioinfo;
1237 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1238 		return (-1);
1239 
1240 	/* Create and build a buffer header for a transfer. */
1241 	bp = (struct buf *)getpbuf(NULL);
1242 	BUF_KERNPROC(bp);
1243 
1244 	AIO_LOCK(ki);
1245 	ki->kaio_count++;
1246 	ki->kaio_buffer_count++;
1247 	lj = aiocbe->lio;
1248 	if (lj)
1249 		lj->lioj_count++;
1250 	AIO_UNLOCK(ki);
1251 
1252 	/*
1253 	 * Get a copy of the kva from the physical buffer.
1254 	 */
1255 	error = 0;
1256 
1257 	bp->b_bcount = cb->aio_nbytes;
1258 	bp->b_bufsize = cb->aio_nbytes;
1259 	bp->b_iodone = aio_physwakeup;
1260 	bp->b_saveaddr = bp->b_data;
1261 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1262 	bp->b_offset = cb->aio_offset;
1263 	bp->b_iooffset = cb->aio_offset;
1264 	bp->b_blkno = btodb(cb->aio_offset);
1265 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1266 
1267 	/*
1268 	 * Bring buffer into kernel space.
1269 	 */
1270 	if (vmapbuf(bp) < 0) {
1271 		error = EFAULT;
1272 		goto doerror;
1273 	}
1274 
1275 	AIO_LOCK(ki);
1276 	aiocbe->bp = bp;
1277 	bp->b_caller1 = (void *)aiocbe;
1278 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1279 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1280 	aiocbe->jobstate = JOBST_JOBQBUF;
1281 	cb->_aiocb_private.status = cb->aio_nbytes;
1282 	AIO_UNLOCK(ki);
1283 
1284 	atomic_add_int(&num_queue_count, 1);
1285 	atomic_add_int(&num_buf_aio, 1);
1286 
1287 	bp->b_error = 0;
1288 
1289 	TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1290 
1291 	/* Perform transfer. */
1292 	dev_strategy(vp->v_rdev, bp);
1293 	return (0);
1294 
1295 doerror:
1296 	AIO_LOCK(ki);
1297 	ki->kaio_count--;
1298 	ki->kaio_buffer_count--;
1299 	if (lj)
1300 		lj->lioj_count--;
1301 	aiocbe->bp = NULL;
1302 	AIO_UNLOCK(ki);
1303 	relpbuf(bp, NULL);
1304 	return (error);
1305 }
1306 
1307 /*
1308  * Wake up aio requests that may be serviceable now.
1309  */
1310 static void
1311 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1312 {
1313 	struct aiocblist *cb, *cbn;
1314 	int opcode;
1315 
1316 	if (sb == &so->so_snd)
1317 		opcode = LIO_WRITE;
1318 	else
1319 		opcode = LIO_READ;
1320 
1321 	SOCKBUF_LOCK(sb);
1322 	sb->sb_flags &= ~SB_AIO;
1323 	mtx_lock(&aio_job_mtx);
1324 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1325 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1326 			if (cb->jobstate != JOBST_JOBQSOCK)
1327 				panic("invalid queue value");
1328 			/* XXX
1329 			 * We don't have actual sockets backend yet,
1330 			 * so we simply move the requests to the generic
1331 			 * file I/O backend.
1332 			 */
1333 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1334 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1335 			aio_kick_nowait(cb->userproc);
1336 		}
1337 	}
1338 	mtx_unlock(&aio_job_mtx);
1339 	SOCKBUF_UNLOCK(sb);
1340 }
1341 
1342 static int
1343 convert_old_sigevent(struct osigevent *osig, struct sigevent *nsig)
1344 {
1345 
1346 	/*
1347 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
1348 	 * supported by AIO with the old sigevent structure.
1349 	 */
1350 	nsig->sigev_notify = osig->sigev_notify;
1351 	switch (nsig->sigev_notify) {
1352 	case SIGEV_NONE:
1353 		break;
1354 	case SIGEV_SIGNAL:
1355 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
1356 		break;
1357 	case SIGEV_KEVENT:
1358 		nsig->sigev_notify_kqueue =
1359 		    osig->__sigev_u.__sigev_notify_kqueue;
1360 		nsig->sigev_value.sival_ptr = osig->sigev_value.sival_ptr;
1361 		break;
1362 	default:
1363 		return (EINVAL);
1364 	}
1365 	return (0);
1366 }
1367 
1368 static int
1369 aiocb_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
1370 {
1371 	struct oaiocb *ojob;
1372 	int error;
1373 
1374 	bzero(kjob, sizeof(struct aiocb));
1375 	error = copyin(ujob, kjob, sizeof(struct oaiocb));
1376 	if (error)
1377 		return (error);
1378 	ojob = (struct oaiocb *)kjob;
1379 	return (convert_old_sigevent(&ojob->aio_sigevent, &kjob->aio_sigevent));
1380 }
1381 
1382 static int
1383 aiocb_copyin(struct aiocb *ujob, struct aiocb *kjob)
1384 {
1385 
1386 	return (copyin(ujob, kjob, sizeof(struct aiocb)));
1387 }
1388 
1389 static long
1390 aiocb_fetch_status(struct aiocb *ujob)
1391 {
1392 
1393 	return (fuword(&ujob->_aiocb_private.status));
1394 }
1395 
1396 static long
1397 aiocb_fetch_error(struct aiocb *ujob)
1398 {
1399 
1400 	return (fuword(&ujob->_aiocb_private.error));
1401 }
1402 
1403 static int
1404 aiocb_store_status(struct aiocb *ujob, long status)
1405 {
1406 
1407 	return (suword(&ujob->_aiocb_private.status, status));
1408 }
1409 
1410 static int
1411 aiocb_store_error(struct aiocb *ujob, long error)
1412 {
1413 
1414 	return (suword(&ujob->_aiocb_private.error, error));
1415 }
1416 
1417 static int
1418 aiocb_store_kernelinfo(struct aiocb *ujob, long jobref)
1419 {
1420 
1421 	return (suword(&ujob->_aiocb_private.kernelinfo, jobref));
1422 }
1423 
1424 static int
1425 aiocb_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
1426 {
1427 
1428 	return (suword(ujobp, (long)ujob));
1429 }
1430 
1431 static struct aiocb_ops aiocb_ops = {
1432 	.copyin = aiocb_copyin,
1433 	.fetch_status = aiocb_fetch_status,
1434 	.fetch_error = aiocb_fetch_error,
1435 	.store_status = aiocb_store_status,
1436 	.store_error = aiocb_store_error,
1437 	.store_kernelinfo = aiocb_store_kernelinfo,
1438 	.store_aiocb = aiocb_store_aiocb,
1439 };
1440 
1441 static struct aiocb_ops aiocb_ops_osigevent = {
1442 	.copyin = aiocb_copyin_old_sigevent,
1443 	.fetch_status = aiocb_fetch_status,
1444 	.fetch_error = aiocb_fetch_error,
1445 	.store_status = aiocb_store_status,
1446 	.store_error = aiocb_store_error,
1447 	.store_kernelinfo = aiocb_store_kernelinfo,
1448 	.store_aiocb = aiocb_store_aiocb,
1449 };
1450 
1451 /*
1452  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1453  * technique is done in this code.
1454  */
1455 int
1456 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1457 	int type, struct aiocb_ops *ops)
1458 {
1459 	struct proc *p = td->td_proc;
1460 	struct file *fp;
1461 	struct socket *so;
1462 	struct aiocblist *aiocbe, *cb;
1463 	struct kaioinfo *ki;
1464 	struct kevent kev;
1465 	struct sockbuf *sb;
1466 	int opcode;
1467 	int error;
1468 	int fd, kqfd;
1469 	int jid;
1470 
1471 	if (p->p_aioinfo == NULL)
1472 		aio_init_aioinfo(p);
1473 
1474 	ki = p->p_aioinfo;
1475 
1476 	ops->store_status(job, -1);
1477 	ops->store_error(job, 0);
1478 	ops->store_kernelinfo(job, -1);
1479 
1480 	if (num_queue_count >= max_queue_count ||
1481 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1482 		ops->store_error(job, EAGAIN);
1483 		return (EAGAIN);
1484 	}
1485 
1486 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1487 	aiocbe->inputcharge = 0;
1488 	aiocbe->outputcharge = 0;
1489 	knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL);
1490 
1491 	error = ops->copyin(job, &aiocbe->uaiocb);
1492 	if (error) {
1493 		ops->store_error(job, error);
1494 		uma_zfree(aiocb_zone, aiocbe);
1495 		return (error);
1496 	}
1497 
1498 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1499 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1500 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1501 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1502 		ops->store_error(job, EINVAL);
1503 		uma_zfree(aiocb_zone, aiocbe);
1504 		return (EINVAL);
1505 	}
1506 
1507 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1508 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1509 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1510 		uma_zfree(aiocb_zone, aiocbe);
1511 		return (EINVAL);
1512 	}
1513 
1514 	ksiginfo_init(&aiocbe->ksi);
1515 
1516 	/* Save userspace address of the job info. */
1517 	aiocbe->uuaiocb = job;
1518 
1519 	/* Get the opcode. */
1520 	if (type != LIO_NOP)
1521 		aiocbe->uaiocb.aio_lio_opcode = type;
1522 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1523 
1524 	/* Fetch the file object for the specified file descriptor. */
1525 	fd = aiocbe->uaiocb.aio_fildes;
1526 	switch (opcode) {
1527 	case LIO_WRITE:
1528 		error = fget_write(td, fd, &fp);
1529 		break;
1530 	case LIO_READ:
1531 		error = fget_read(td, fd, &fp);
1532 		break;
1533 	default:
1534 		error = fget(td, fd, &fp);
1535 	}
1536 	if (error) {
1537 		uma_zfree(aiocb_zone, aiocbe);
1538 		ops->store_error(job, error);
1539 		return (error);
1540 	}
1541 
1542 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1543 		error = EINVAL;
1544 		goto aqueue_fail;
1545 	}
1546 
1547 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1548 		error = EINVAL;
1549 		goto aqueue_fail;
1550 	}
1551 
1552 	aiocbe->fd_file = fp;
1553 
1554 	mtx_lock(&aio_job_mtx);
1555 	jid = jobrefid++;
1556 	aiocbe->seqno = jobseqno++;
1557 	mtx_unlock(&aio_job_mtx);
1558 	error = ops->store_kernelinfo(job, jid);
1559 	if (error) {
1560 		error = EINVAL;
1561 		goto aqueue_fail;
1562 	}
1563 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1564 
1565 	if (opcode == LIO_NOP) {
1566 		fdrop(fp, td);
1567 		uma_zfree(aiocb_zone, aiocbe);
1568 		return (0);
1569 	}
1570 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1571 	    (opcode != LIO_SYNC)) {
1572 		error = EINVAL;
1573 		goto aqueue_fail;
1574 	}
1575 
1576 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1577 		goto no_kqueue;
1578 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1579 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1580 	kev.filter = EVFILT_AIO;
1581 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1582 	kev.data = (intptr_t)aiocbe;
1583 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1584 	error = kqfd_register(kqfd, &kev, td, 1);
1585 aqueue_fail:
1586 	if (error) {
1587 		fdrop(fp, td);
1588 		uma_zfree(aiocb_zone, aiocbe);
1589 		ops->store_error(job, error);
1590 		goto done;
1591 	}
1592 no_kqueue:
1593 
1594 	ops->store_error(job, EINPROGRESS);
1595 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1596 	aiocbe->userproc = p;
1597 	aiocbe->cred = crhold(td->td_ucred);
1598 	aiocbe->jobflags = 0;
1599 	aiocbe->lio = lj;
1600 
1601 	if (opcode == LIO_SYNC)
1602 		goto queueit;
1603 
1604 	if (fp->f_type == DTYPE_SOCKET) {
1605 		/*
1606 		 * Alternate queueing for socket ops: Reach down into the
1607 		 * descriptor to get the socket data.  Then check to see if the
1608 		 * socket is ready to be read or written (based on the requested
1609 		 * operation).
1610 		 *
1611 		 * If it is not ready for io, then queue the aiocbe on the
1612 		 * socket, and set the flags so we get a call when sbnotify()
1613 		 * happens.
1614 		 *
1615 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1616 		 * and unlock the snd sockbuf for no reason.
1617 		 */
1618 		so = fp->f_data;
1619 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1620 		SOCKBUF_LOCK(sb);
1621 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1622 		    LIO_WRITE) && (!sowriteable(so)))) {
1623 			sb->sb_flags |= SB_AIO;
1624 
1625 			mtx_lock(&aio_job_mtx);
1626 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1627 			mtx_unlock(&aio_job_mtx);
1628 
1629 			AIO_LOCK(ki);
1630 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1631 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1632 			aiocbe->jobstate = JOBST_JOBQSOCK;
1633 			ki->kaio_count++;
1634 			if (lj)
1635 				lj->lioj_count++;
1636 			AIO_UNLOCK(ki);
1637 			SOCKBUF_UNLOCK(sb);
1638 			atomic_add_int(&num_queue_count, 1);
1639 			error = 0;
1640 			goto done;
1641 		}
1642 		SOCKBUF_UNLOCK(sb);
1643 	}
1644 
1645 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1646 		goto done;
1647 #if 0
1648 	if (error > 0) {
1649 		aiocbe->uaiocb._aiocb_private.error = error;
1650 		ops->store_error(job, error);
1651 		goto done;
1652 	}
1653 #endif
1654 queueit:
1655 	/* No buffer for daemon I/O. */
1656 	aiocbe->bp = NULL;
1657 	atomic_add_int(&num_queue_count, 1);
1658 
1659 	AIO_LOCK(ki);
1660 	ki->kaio_count++;
1661 	if (lj)
1662 		lj->lioj_count++;
1663 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1664 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1665 	if (opcode == LIO_SYNC) {
1666 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1667 			if (cb->fd_file == aiocbe->fd_file &&
1668 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1669 			    cb->seqno < aiocbe->seqno) {
1670 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1671 				aiocbe->pending++;
1672 			}
1673 		}
1674 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1675 			if (cb->fd_file == aiocbe->fd_file &&
1676 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1677 			    cb->seqno < aiocbe->seqno) {
1678 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1679 				aiocbe->pending++;
1680 			}
1681 		}
1682 		if (aiocbe->pending != 0) {
1683 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1684 			aiocbe->jobstate = JOBST_JOBQSYNC;
1685 			AIO_UNLOCK(ki);
1686 			goto done;
1687 		}
1688 	}
1689 	mtx_lock(&aio_job_mtx);
1690 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1691 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1692 	aio_kick_nowait(p);
1693 	mtx_unlock(&aio_job_mtx);
1694 	AIO_UNLOCK(ki);
1695 	error = 0;
1696 done:
1697 	return (error);
1698 }
1699 
1700 static void
1701 aio_kick_nowait(struct proc *userp)
1702 {
1703 	struct kaioinfo *ki = userp->p_aioinfo;
1704 	struct aiothreadlist *aiop;
1705 
1706 	mtx_assert(&aio_job_mtx, MA_OWNED);
1707 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1708 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1709 		aiop->aiothreadflags &= ~AIOP_FREE;
1710 		wakeup(aiop->aiothread);
1711 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1712 	    ((ki->kaio_active_count + num_aio_resv_start) <
1713 	    ki->kaio_maxactive_count)) {
1714 		taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1715 	}
1716 }
1717 
1718 static int
1719 aio_kick(struct proc *userp)
1720 {
1721 	struct kaioinfo *ki = userp->p_aioinfo;
1722 	struct aiothreadlist *aiop;
1723 	int error, ret = 0;
1724 
1725 	mtx_assert(&aio_job_mtx, MA_OWNED);
1726 retryproc:
1727 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1728 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1729 		aiop->aiothreadflags &= ~AIOP_FREE;
1730 		wakeup(aiop->aiothread);
1731 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1732 	    ((ki->kaio_active_count + num_aio_resv_start) <
1733 	    ki->kaio_maxactive_count)) {
1734 		num_aio_resv_start++;
1735 		mtx_unlock(&aio_job_mtx);
1736 		error = aio_newproc(&num_aio_resv_start);
1737 		mtx_lock(&aio_job_mtx);
1738 		if (error) {
1739 			num_aio_resv_start--;
1740 			goto retryproc;
1741 		}
1742 	} else {
1743 		ret = -1;
1744 	}
1745 	return (ret);
1746 }
1747 
1748 static void
1749 aio_kick_helper(void *context, int pending)
1750 {
1751 	struct proc *userp = context;
1752 
1753 	mtx_lock(&aio_job_mtx);
1754 	while (--pending >= 0) {
1755 		if (aio_kick(userp))
1756 			break;
1757 	}
1758 	mtx_unlock(&aio_job_mtx);
1759 }
1760 
1761 /*
1762  * Support the aio_return system call, as a side-effect, kernel resources are
1763  * released.
1764  */
1765 static int
1766 kern_aio_return(struct thread *td, struct aiocb *uaiocb, struct aiocb_ops *ops)
1767 {
1768 	struct proc *p = td->td_proc;
1769 	struct aiocblist *cb;
1770 	struct kaioinfo *ki;
1771 	int status, error;
1772 
1773 	ki = p->p_aioinfo;
1774 	if (ki == NULL)
1775 		return (EINVAL);
1776 	AIO_LOCK(ki);
1777 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1778 		if (cb->uuaiocb == uaiocb)
1779 			break;
1780 	}
1781 	if (cb != NULL) {
1782 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1783 		status = cb->uaiocb._aiocb_private.status;
1784 		error = cb->uaiocb._aiocb_private.error;
1785 		td->td_retval[0] = status;
1786 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1787 			td->td_ru.ru_oublock += cb->outputcharge;
1788 			cb->outputcharge = 0;
1789 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1790 			td->td_ru.ru_inblock += cb->inputcharge;
1791 			cb->inputcharge = 0;
1792 		}
1793 		aio_free_entry(cb);
1794 		AIO_UNLOCK(ki);
1795 		ops->store_error(uaiocb, error);
1796 		ops->store_status(uaiocb, status);
1797 	} else {
1798 		error = EINVAL;
1799 		AIO_UNLOCK(ki);
1800 	}
1801 	return (error);
1802 }
1803 
1804 int
1805 aio_return(struct thread *td, struct aio_return_args *uap)
1806 {
1807 
1808 	return (kern_aio_return(td, uap->aiocbp, &aiocb_ops));
1809 }
1810 
1811 /*
1812  * Allow a process to wakeup when any of the I/O requests are completed.
1813  */
1814 static int
1815 kern_aio_suspend(struct thread *td, int njoblist, struct aiocb **ujoblist,
1816     struct timespec *ts)
1817 {
1818 	struct proc *p = td->td_proc;
1819 	struct timeval atv;
1820 	struct kaioinfo *ki;
1821 	struct aiocblist *cb, *cbfirst;
1822 	int error, i, timo;
1823 
1824 	timo = 0;
1825 	if (ts) {
1826 		if (ts->tv_nsec < 0 || ts->tv_nsec >= 1000000000)
1827 			return (EINVAL);
1828 
1829 		TIMESPEC_TO_TIMEVAL(&atv, ts);
1830 		if (itimerfix(&atv))
1831 			return (EINVAL);
1832 		timo = tvtohz(&atv);
1833 	}
1834 
1835 	ki = p->p_aioinfo;
1836 	if (ki == NULL)
1837 		return (EAGAIN);
1838 
1839 	if (njoblist == 0)
1840 		return (0);
1841 
1842 	AIO_LOCK(ki);
1843 	for (;;) {
1844 		cbfirst = NULL;
1845 		error = 0;
1846 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1847 			for (i = 0; i < njoblist; i++) {
1848 				if (cb->uuaiocb == ujoblist[i]) {
1849 					if (cbfirst == NULL)
1850 						cbfirst = cb;
1851 					if (cb->jobstate == JOBST_JOBFINISHED)
1852 						goto RETURN;
1853 				}
1854 			}
1855 		}
1856 		/* All tasks were finished. */
1857 		if (cbfirst == NULL)
1858 			break;
1859 
1860 		ki->kaio_flags |= KAIO_WAKEUP;
1861 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1862 		    "aiospn", timo);
1863 		if (error == ERESTART)
1864 			error = EINTR;
1865 		if (error)
1866 			break;
1867 	}
1868 RETURN:
1869 	AIO_UNLOCK(ki);
1870 	return (error);
1871 }
1872 
1873 int
1874 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1875 {
1876 	struct timespec ts, *tsp;
1877 	struct aiocb **ujoblist;
1878 	int error;
1879 
1880 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1881 		return (EINVAL);
1882 
1883 	if (uap->timeout) {
1884 		/* Get timespec struct. */
1885 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1886 			return (error);
1887 		tsp = &ts;
1888 	} else
1889 		tsp = NULL;
1890 
1891 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1892 	error = copyin(uap->aiocbp, ujoblist, uap->nent * sizeof(ujoblist[0]));
1893 	if (error == 0)
1894 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
1895 	uma_zfree(aiol_zone, ujoblist);
1896 	return (error);
1897 }
1898 
1899 /*
1900  * aio_cancel cancels any non-physio aio operations not currently in
1901  * progress.
1902  */
1903 int
1904 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1905 {
1906 	struct proc *p = td->td_proc;
1907 	struct kaioinfo *ki;
1908 	struct aiocblist *cbe, *cbn;
1909 	struct file *fp;
1910 	struct socket *so;
1911 	int error;
1912 	int remove;
1913 	int cancelled = 0;
1914 	int notcancelled = 0;
1915 	struct vnode *vp;
1916 
1917 	/* Lookup file object. */
1918 	error = fget(td, uap->fd, &fp);
1919 	if (error)
1920 		return (error);
1921 
1922 	ki = p->p_aioinfo;
1923 	if (ki == NULL)
1924 		goto done;
1925 
1926 	if (fp->f_type == DTYPE_VNODE) {
1927 		vp = fp->f_vnode;
1928 		if (vn_isdisk(vp, &error)) {
1929 			fdrop(fp, td);
1930 			td->td_retval[0] = AIO_NOTCANCELED;
1931 			return (0);
1932 		}
1933 	}
1934 
1935 	AIO_LOCK(ki);
1936 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1937 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1938 		    ((uap->aiocbp == NULL) ||
1939 		     (uap->aiocbp == cbe->uuaiocb))) {
1940 			remove = 0;
1941 
1942 			mtx_lock(&aio_job_mtx);
1943 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1944 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1945 				remove = 1;
1946 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
1947 				MPASS(fp->f_type == DTYPE_SOCKET);
1948 				so = fp->f_data;
1949 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1950 				remove = 1;
1951 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
1952 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1953 				remove = 1;
1954 			}
1955 			mtx_unlock(&aio_job_mtx);
1956 
1957 			if (remove) {
1958 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1959 				cbe->uaiocb._aiocb_private.status = -1;
1960 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1961 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
1962 				cancelled++;
1963 			} else {
1964 				notcancelled++;
1965 			}
1966 			if (uap->aiocbp != NULL)
1967 				break;
1968 		}
1969 	}
1970 	AIO_UNLOCK(ki);
1971 
1972 done:
1973 	fdrop(fp, td);
1974 
1975 	if (uap->aiocbp != NULL) {
1976 		if (cancelled) {
1977 			td->td_retval[0] = AIO_CANCELED;
1978 			return (0);
1979 		}
1980 	}
1981 
1982 	if (notcancelled) {
1983 		td->td_retval[0] = AIO_NOTCANCELED;
1984 		return (0);
1985 	}
1986 
1987 	if (cancelled) {
1988 		td->td_retval[0] = AIO_CANCELED;
1989 		return (0);
1990 	}
1991 
1992 	td->td_retval[0] = AIO_ALLDONE;
1993 
1994 	return (0);
1995 }
1996 
1997 /*
1998  * aio_error is implemented in the kernel level for compatibility purposes
1999  * only.  For a user mode async implementation, it would be best to do it in
2000  * a userland subroutine.
2001  */
2002 static int
2003 kern_aio_error(struct thread *td, struct aiocb *aiocbp, struct aiocb_ops *ops)
2004 {
2005 	struct proc *p = td->td_proc;
2006 	struct aiocblist *cb;
2007 	struct kaioinfo *ki;
2008 	int status;
2009 
2010 	ki = p->p_aioinfo;
2011 	if (ki == NULL) {
2012 		td->td_retval[0] = EINVAL;
2013 		return (0);
2014 	}
2015 
2016 	AIO_LOCK(ki);
2017 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
2018 		if (cb->uuaiocb == aiocbp) {
2019 			if (cb->jobstate == JOBST_JOBFINISHED)
2020 				td->td_retval[0] =
2021 					cb->uaiocb._aiocb_private.error;
2022 			else
2023 				td->td_retval[0] = EINPROGRESS;
2024 			AIO_UNLOCK(ki);
2025 			return (0);
2026 		}
2027 	}
2028 	AIO_UNLOCK(ki);
2029 
2030 	/*
2031 	 * Hack for failure of aio_aqueue.
2032 	 */
2033 	status = ops->fetch_status(aiocbp);
2034 	if (status == -1) {
2035 		td->td_retval[0] = ops->fetch_error(aiocbp);
2036 		return (0);
2037 	}
2038 
2039 	td->td_retval[0] = EINVAL;
2040 	return (0);
2041 }
2042 
2043 int
2044 aio_error(struct thread *td, struct aio_error_args *uap)
2045 {
2046 
2047 	return (kern_aio_error(td, uap->aiocbp, &aiocb_ops));
2048 }
2049 
2050 /* syscall - asynchronous read from a file (REALTIME) */
2051 int
2052 oaio_read(struct thread *td, struct oaio_read_args *uap)
2053 {
2054 
2055 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2056 	    &aiocb_ops_osigevent));
2057 }
2058 
2059 int
2060 aio_read(struct thread *td, struct aio_read_args *uap)
2061 {
2062 
2063 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, &aiocb_ops));
2064 }
2065 
2066 /* syscall - asynchronous write to a file (REALTIME) */
2067 int
2068 oaio_write(struct thread *td, struct oaio_write_args *uap)
2069 {
2070 
2071 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2072 	    &aiocb_ops_osigevent));
2073 }
2074 
2075 int
2076 aio_write(struct thread *td, struct aio_write_args *uap)
2077 {
2078 
2079 	return (aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, &aiocb_ops));
2080 }
2081 
2082 static int
2083 kern_lio_listio(struct thread *td, int mode, struct aiocb * const *uacb_list,
2084     struct aiocb **acb_list, int nent, struct sigevent *sig,
2085     struct aiocb_ops *ops)
2086 {
2087 	struct proc *p = td->td_proc;
2088 	struct aiocb *iocb;
2089 	struct kaioinfo *ki;
2090 	struct aioliojob *lj;
2091 	struct kevent kev;
2092 	int error;
2093 	int nerror;
2094 	int i;
2095 
2096 	if ((mode != LIO_NOWAIT) && (mode != LIO_WAIT))
2097 		return (EINVAL);
2098 
2099 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2100 		return (EINVAL);
2101 
2102 	if (p->p_aioinfo == NULL)
2103 		aio_init_aioinfo(p);
2104 
2105 	ki = p->p_aioinfo;
2106 
2107 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
2108 	lj->lioj_flags = 0;
2109 	lj->lioj_count = 0;
2110 	lj->lioj_finished_count = 0;
2111 	knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL);
2112 	ksiginfo_init(&lj->lioj_ksi);
2113 
2114 	/*
2115 	 * Setup signal.
2116 	 */
2117 	if (sig && (mode == LIO_NOWAIT)) {
2118 		bcopy(sig, &lj->lioj_signal, sizeof(lj->lioj_signal));
2119 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2120 			/* Assume only new style KEVENT */
2121 			kev.filter = EVFILT_LIO;
2122 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2123 			kev.ident = (uintptr_t)uacb_list; /* something unique */
2124 			kev.data = (intptr_t)lj;
2125 			/* pass user defined sigval data */
2126 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2127 			error = kqfd_register(
2128 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2129 			if (error) {
2130 				uma_zfree(aiolio_zone, lj);
2131 				return (error);
2132 			}
2133 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2134 			;
2135 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2136 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2137 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2138 					uma_zfree(aiolio_zone, lj);
2139 					return EINVAL;
2140 				}
2141 				lj->lioj_flags |= LIOJ_SIGNAL;
2142 		} else {
2143 			uma_zfree(aiolio_zone, lj);
2144 			return EINVAL;
2145 		}
2146 	}
2147 
2148 	AIO_LOCK(ki);
2149 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2150 	/*
2151 	 * Add extra aiocb count to avoid the lio to be freed
2152 	 * by other threads doing aio_waitcomplete or aio_return,
2153 	 * and prevent event from being sent until we have queued
2154 	 * all tasks.
2155 	 */
2156 	lj->lioj_count = 1;
2157 	AIO_UNLOCK(ki);
2158 
2159 	/*
2160 	 * Get pointers to the list of I/O requests.
2161 	 */
2162 	nerror = 0;
2163 	for (i = 0; i < nent; i++) {
2164 		iocb = acb_list[i];
2165 		if (iocb != NULL) {
2166 			error = aio_aqueue(td, iocb, lj, LIO_NOP, ops);
2167 			if (error != 0)
2168 				nerror++;
2169 		}
2170 	}
2171 
2172 	error = 0;
2173 	AIO_LOCK(ki);
2174 	if (mode == LIO_WAIT) {
2175 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2176 			ki->kaio_flags |= KAIO_WAKEUP;
2177 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2178 			    PRIBIO | PCATCH, "aiospn", 0);
2179 			if (error == ERESTART)
2180 				error = EINTR;
2181 			if (error)
2182 				break;
2183 		}
2184 	} else {
2185 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2186 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2187 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2188 				KNOTE_LOCKED(&lj->klist, 1);
2189 			}
2190 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2191 			    == LIOJ_SIGNAL
2192 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2193 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2194 				aio_sendsig(p, &lj->lioj_signal,
2195 					    &lj->lioj_ksi);
2196 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2197 			}
2198 		}
2199 	}
2200 	lj->lioj_count--;
2201 	if (lj->lioj_count == 0) {
2202 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2203 		knlist_delete(&lj->klist, curthread, 1);
2204 		PROC_LOCK(p);
2205 		sigqueue_take(&lj->lioj_ksi);
2206 		PROC_UNLOCK(p);
2207 		AIO_UNLOCK(ki);
2208 		uma_zfree(aiolio_zone, lj);
2209 	} else
2210 		AIO_UNLOCK(ki);
2211 
2212 	if (nerror)
2213 		return (EIO);
2214 	return (error);
2215 }
2216 
2217 /* syscall - list directed I/O (REALTIME) */
2218 int
2219 olio_listio(struct thread *td, struct olio_listio_args *uap)
2220 {
2221 	struct aiocb **acb_list;
2222 	struct sigevent *sigp, sig;
2223 	struct osigevent osig;
2224 	int error, nent;
2225 
2226 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2227 		return (EINVAL);
2228 
2229 	nent = uap->nent;
2230 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2231 		return (EINVAL);
2232 
2233 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2234 		error = copyin(uap->sig, &osig, sizeof(osig));
2235 		if (error)
2236 			return (error);
2237 		error = convert_old_sigevent(&osig, &sig);
2238 		if (error)
2239 			return (error);
2240 		sigp = &sig;
2241 	} else
2242 		sigp = NULL;
2243 
2244 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2245 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2246 	if (error == 0)
2247 		error = kern_lio_listio(td, uap->mode,
2248 		    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2249 		    &aiocb_ops_osigevent);
2250 	free(acb_list, M_LIO);
2251 	return (error);
2252 }
2253 
2254 /* syscall - list directed I/O (REALTIME) */
2255 int
2256 lio_listio(struct thread *td, struct lio_listio_args *uap)
2257 {
2258 	struct aiocb **acb_list;
2259 	struct sigevent *sigp, sig;
2260 	int error, nent;
2261 
2262 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2263 		return (EINVAL);
2264 
2265 	nent = uap->nent;
2266 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2267 		return (EINVAL);
2268 
2269 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2270 		error = copyin(uap->sig, &sig, sizeof(sig));
2271 		if (error)
2272 			return (error);
2273 		sigp = &sig;
2274 	} else
2275 		sigp = NULL;
2276 
2277 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2278 	error = copyin(uap->acb_list, acb_list, nent * sizeof(acb_list[0]));
2279 	if (error == 0)
2280 		error = kern_lio_listio(td, uap->mode, uap->acb_list, acb_list,
2281 		    nent, sigp, &aiocb_ops);
2282 	free(acb_list, M_LIO);
2283 	return (error);
2284 }
2285 
2286 /*
2287  * Called from interrupt thread for physio, we should return as fast
2288  * as possible, so we schedule a biohelper task.
2289  */
2290 static void
2291 aio_physwakeup(struct buf *bp)
2292 {
2293 	struct aiocblist *aiocbe;
2294 
2295 	aiocbe = (struct aiocblist *)bp->b_caller1;
2296 	taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2297 }
2298 
2299 /*
2300  * Task routine to perform heavy tasks, process wakeup, and signals.
2301  */
2302 static void
2303 biohelper(void *context, int pending)
2304 {
2305 	struct aiocblist *aiocbe = context;
2306 	struct buf *bp;
2307 	struct proc *userp;
2308 	struct kaioinfo *ki;
2309 	int nblks;
2310 
2311 	bp = aiocbe->bp;
2312 	userp = aiocbe->userproc;
2313 	ki = userp->p_aioinfo;
2314 	AIO_LOCK(ki);
2315 	aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2316 	aiocbe->uaiocb._aiocb_private.error = 0;
2317 	if (bp->b_ioflags & BIO_ERROR)
2318 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2319 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2320 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2321 		aiocbe->outputcharge += nblks;
2322 	else
2323 		aiocbe->inputcharge += nblks;
2324 	aiocbe->bp = NULL;
2325 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2326 	ki->kaio_buffer_count--;
2327 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2328 	AIO_UNLOCK(ki);
2329 
2330 	/* Release mapping into kernel space. */
2331 	vunmapbuf(bp);
2332 	relpbuf(bp, NULL);
2333 	atomic_subtract_int(&num_buf_aio, 1);
2334 }
2335 
2336 /* syscall - wait for the next completion of an aio request */
2337 static int
2338 kern_aio_waitcomplete(struct thread *td, struct aiocb **aiocbp,
2339     struct timespec *ts, struct aiocb_ops *ops)
2340 {
2341 	struct proc *p = td->td_proc;
2342 	struct timeval atv;
2343 	struct kaioinfo *ki;
2344 	struct aiocblist *cb;
2345 	struct aiocb *uuaiocb;
2346 	int error, status, timo;
2347 
2348 	ops->store_aiocb(aiocbp, NULL);
2349 
2350 	timo = 0;
2351 	if (ts) {
2352 		if ((ts->tv_nsec < 0) || (ts->tv_nsec >= 1000000000))
2353 			return (EINVAL);
2354 
2355 		TIMESPEC_TO_TIMEVAL(&atv, ts);
2356 		if (itimerfix(&atv))
2357 			return (EINVAL);
2358 		timo = tvtohz(&atv);
2359 	}
2360 
2361 	if (p->p_aioinfo == NULL)
2362 		aio_init_aioinfo(p);
2363 	ki = p->p_aioinfo;
2364 
2365 	error = 0;
2366 	cb = NULL;
2367 	AIO_LOCK(ki);
2368 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2369 		ki->kaio_flags |= KAIO_WAKEUP;
2370 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2371 		    "aiowc", timo);
2372 		if (timo && error == ERESTART)
2373 			error = EINTR;
2374 		if (error)
2375 			break;
2376 	}
2377 
2378 	if (cb != NULL) {
2379 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2380 		uuaiocb = cb->uuaiocb;
2381 		status = cb->uaiocb._aiocb_private.status;
2382 		error = cb->uaiocb._aiocb_private.error;
2383 		td->td_retval[0] = status;
2384 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2385 			td->td_ru.ru_oublock += cb->outputcharge;
2386 			cb->outputcharge = 0;
2387 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2388 			td->td_ru.ru_inblock += cb->inputcharge;
2389 			cb->inputcharge = 0;
2390 		}
2391 		aio_free_entry(cb);
2392 		AIO_UNLOCK(ki);
2393 		ops->store_aiocb(aiocbp, uuaiocb);
2394 		ops->store_error(uuaiocb, error);
2395 		ops->store_status(uuaiocb, status);
2396 	} else
2397 		AIO_UNLOCK(ki);
2398 
2399 	return (error);
2400 }
2401 
2402 int
2403 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2404 {
2405 	struct timespec ts, *tsp;
2406 	int error;
2407 
2408 	if (uap->timeout) {
2409 		/* Get timespec struct. */
2410 		error = copyin(uap->timeout, &ts, sizeof(ts));
2411 		if (error)
2412 			return (error);
2413 		tsp = &ts;
2414 	} else
2415 		tsp = NULL;
2416 
2417 	return (kern_aio_waitcomplete(td, uap->aiocbp, tsp, &aiocb_ops));
2418 }
2419 
2420 static int
2421 kern_aio_fsync(struct thread *td, int op, struct aiocb *aiocbp,
2422     struct aiocb_ops *ops)
2423 {
2424 	struct proc *p = td->td_proc;
2425 	struct kaioinfo *ki;
2426 
2427 	if (op != O_SYNC) /* XXX lack of O_DSYNC */
2428 		return (EINVAL);
2429 	ki = p->p_aioinfo;
2430 	if (ki == NULL)
2431 		aio_init_aioinfo(p);
2432 	return (aio_aqueue(td, aiocbp, NULL, LIO_SYNC, ops));
2433 }
2434 
2435 int
2436 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2437 {
2438 
2439 	return (kern_aio_fsync(td, uap->op, uap->aiocbp, &aiocb_ops));
2440 }
2441 
2442 /* kqueue attach function */
2443 static int
2444 filt_aioattach(struct knote *kn)
2445 {
2446 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2447 
2448 	/*
2449 	 * The aiocbe pointer must be validated before using it, so
2450 	 * registration is restricted to the kernel; the user cannot
2451 	 * set EV_FLAG1.
2452 	 */
2453 	if ((kn->kn_flags & EV_FLAG1) == 0)
2454 		return (EPERM);
2455 	kn->kn_ptr.p_aio = aiocbe;
2456 	kn->kn_flags &= ~EV_FLAG1;
2457 
2458 	knlist_add(&aiocbe->klist, kn, 0);
2459 
2460 	return (0);
2461 }
2462 
2463 /* kqueue detach function */
2464 static void
2465 filt_aiodetach(struct knote *kn)
2466 {
2467 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2468 
2469 	if (!knlist_empty(&aiocbe->klist))
2470 		knlist_remove(&aiocbe->klist, kn, 0);
2471 }
2472 
2473 /* kqueue filter function */
2474 /*ARGSUSED*/
2475 static int
2476 filt_aio(struct knote *kn, long hint)
2477 {
2478 	struct aiocblist *aiocbe = kn->kn_ptr.p_aio;
2479 
2480 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2481 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2482 		return (0);
2483 	kn->kn_flags |= EV_EOF;
2484 	return (1);
2485 }
2486 
2487 /* kqueue attach function */
2488 static int
2489 filt_lioattach(struct knote *kn)
2490 {
2491 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2492 
2493 	/*
2494 	 * The aioliojob pointer must be validated before using it, so
2495 	 * registration is restricted to the kernel; the user cannot
2496 	 * set EV_FLAG1.
2497 	 */
2498 	if ((kn->kn_flags & EV_FLAG1) == 0)
2499 		return (EPERM);
2500 	kn->kn_ptr.p_lio = lj;
2501 	kn->kn_flags &= ~EV_FLAG1;
2502 
2503 	knlist_add(&lj->klist, kn, 0);
2504 
2505 	return (0);
2506 }
2507 
2508 /* kqueue detach function */
2509 static void
2510 filt_liodetach(struct knote *kn)
2511 {
2512 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2513 
2514 	if (!knlist_empty(&lj->klist))
2515 		knlist_remove(&lj->klist, kn, 0);
2516 }
2517 
2518 /* kqueue filter function */
2519 /*ARGSUSED*/
2520 static int
2521 filt_lio(struct knote *kn, long hint)
2522 {
2523 	struct aioliojob * lj = kn->kn_ptr.p_lio;
2524 
2525 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2526 }
2527 
2528 #ifdef COMPAT_IA32
2529 #include <sys/mount.h>
2530 #include <sys/socket.h>
2531 #include <compat/freebsd32/freebsd32.h>
2532 #include <compat/freebsd32/freebsd32_proto.h>
2533 #include <compat/freebsd32/freebsd32_signal.h>
2534 #include <compat/freebsd32/freebsd32_syscall.h>
2535 #include <compat/freebsd32/freebsd32_util.h>
2536 
2537 struct __aiocb_private32 {
2538 	int32_t	status;
2539 	int32_t	error;
2540 	uint32_t kernelinfo;
2541 };
2542 
2543 typedef struct oaiocb32 {
2544 	int	aio_fildes;		/* File descriptor */
2545 	uint64_t aio_offset __packed;	/* File offset for I/O */
2546 	uint32_t aio_buf;		/* I/O buffer in process space */
2547 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2548 	struct	osigevent32 aio_sigevent; /* Signal to deliver */
2549 	int	aio_lio_opcode;		/* LIO opcode */
2550 	int	aio_reqprio;		/* Request priority -- ignored */
2551 	struct	__aiocb_private32 _aiocb_private;
2552 } oaiocb32_t;
2553 
2554 typedef struct aiocb32 {
2555 	int32_t	aio_fildes;		/* File descriptor */
2556 	uint64_t aio_offset __packed;	/* File offset for I/O */
2557 	uint32_t aio_buf;		/* I/O buffer in process space */
2558 	uint32_t aio_nbytes;		/* Number of bytes for I/O */
2559 	int	__spare__[2];
2560 	uint32_t __spare2__;
2561 	int	aio_lio_opcode;		/* LIO opcode */
2562 	int	aio_reqprio;		/* Request priority -- ignored */
2563 	struct __aiocb_private32 _aiocb_private;
2564 	struct sigevent32 aio_sigevent;	/* Signal to deliver */
2565 } aiocb32_t;
2566 
2567 static int
2568 convert_old_sigevent32(struct osigevent32 *osig, struct sigevent *nsig)
2569 {
2570 
2571 	/*
2572 	 * Only SIGEV_NONE, SIGEV_SIGNAL, and SIGEV_KEVENT are
2573 	 * supported by AIO with the old sigevent structure.
2574 	 */
2575 	CP(*osig, *nsig, sigev_notify);
2576 	switch (nsig->sigev_notify) {
2577 	case SIGEV_NONE:
2578 		break;
2579 	case SIGEV_SIGNAL:
2580 		nsig->sigev_signo = osig->__sigev_u.__sigev_signo;
2581 		break;
2582 	case SIGEV_KEVENT:
2583 		nsig->sigev_notify_kqueue =
2584 		    osig->__sigev_u.__sigev_notify_kqueue;
2585 		PTRIN_CP(*osig, *nsig, sigev_value.sival_ptr);
2586 		break;
2587 	default:
2588 		return (EINVAL);
2589 	}
2590 	return (0);
2591 }
2592 
2593 static int
2594 aiocb32_copyin_old_sigevent(struct aiocb *ujob, struct aiocb *kjob)
2595 {
2596 	struct oaiocb32 job32;
2597 	int error;
2598 
2599 	bzero(kjob, sizeof(struct aiocb));
2600 	error = copyin(ujob, &job32, sizeof(job32));
2601 	if (error)
2602 		return (error);
2603 
2604 	CP(job32, *kjob, aio_fildes);
2605 	CP(job32, *kjob, aio_offset);
2606 	PTRIN_CP(job32, *kjob, aio_buf);
2607 	CP(job32, *kjob, aio_nbytes);
2608 	CP(job32, *kjob, aio_lio_opcode);
2609 	CP(job32, *kjob, aio_reqprio);
2610 	CP(job32, *kjob, _aiocb_private.status);
2611 	CP(job32, *kjob, _aiocb_private.error);
2612 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2613 	return (convert_old_sigevent32(&job32.aio_sigevent,
2614 	    &kjob->aio_sigevent));
2615 }
2616 
2617 static int
2618 convert_sigevent32(struct sigevent32 *sig32, struct sigevent *sig)
2619 {
2620 
2621 	CP(*sig32, *sig, sigev_notify);
2622 	switch (sig->sigev_notify) {
2623 	case SIGEV_NONE:
2624 		break;
2625 	case SIGEV_THREAD_ID:
2626 		CP(*sig32, *sig, sigev_notify_thread_id);
2627 		/* FALLTHROUGH */
2628 	case SIGEV_SIGNAL:
2629 		CP(*sig32, *sig, sigev_signo);
2630 		break;
2631 	case SIGEV_KEVENT:
2632 		CP(*sig32, *sig, sigev_notify_kqueue);
2633 		PTRIN_CP(*sig32, *sig, sigev_value.sival_ptr);
2634 		break;
2635 	default:
2636 		return (EINVAL);
2637 	}
2638 	return (0);
2639 }
2640 
2641 static int
2642 aiocb32_copyin(struct aiocb *ujob, struct aiocb *kjob)
2643 {
2644 	struct aiocb32 job32;
2645 	int error;
2646 
2647 	error = copyin(ujob, &job32, sizeof(job32));
2648 	if (error)
2649 		return (error);
2650 	CP(job32, *kjob, aio_fildes);
2651 	CP(job32, *kjob, aio_offset);
2652 	PTRIN_CP(job32, *kjob, aio_buf);
2653 	CP(job32, *kjob, aio_nbytes);
2654 	CP(job32, *kjob, aio_lio_opcode);
2655 	CP(job32, *kjob, aio_reqprio);
2656 	CP(job32, *kjob, _aiocb_private.status);
2657 	CP(job32, *kjob, _aiocb_private.error);
2658 	PTRIN_CP(job32, *kjob, _aiocb_private.kernelinfo);
2659 	return (convert_sigevent32(&job32.aio_sigevent, &kjob->aio_sigevent));
2660 }
2661 
2662 static long
2663 aiocb32_fetch_status(struct aiocb *ujob)
2664 {
2665 	struct aiocb32 *ujob32;
2666 
2667 	ujob32 = (struct aiocb32 *)ujob;
2668 	return (fuword32(&ujob32->_aiocb_private.status));
2669 }
2670 
2671 static long
2672 aiocb32_fetch_error(struct aiocb *ujob)
2673 {
2674 	struct aiocb32 *ujob32;
2675 
2676 	ujob32 = (struct aiocb32 *)ujob;
2677 	return (fuword32(&ujob32->_aiocb_private.error));
2678 }
2679 
2680 static int
2681 aiocb32_store_status(struct aiocb *ujob, long status)
2682 {
2683 	struct aiocb32 *ujob32;
2684 
2685 	ujob32 = (struct aiocb32 *)ujob;
2686 	return (suword32(&ujob32->_aiocb_private.status, status));
2687 }
2688 
2689 static int
2690 aiocb32_store_error(struct aiocb *ujob, long error)
2691 {
2692 	struct aiocb32 *ujob32;
2693 
2694 	ujob32 = (struct aiocb32 *)ujob;
2695 	return (suword32(&ujob32->_aiocb_private.error, error));
2696 }
2697 
2698 static int
2699 aiocb32_store_kernelinfo(struct aiocb *ujob, long jobref)
2700 {
2701 	struct aiocb32 *ujob32;
2702 
2703 	ujob32 = (struct aiocb32 *)ujob;
2704 	return (suword32(&ujob32->_aiocb_private.kernelinfo, jobref));
2705 }
2706 
2707 static int
2708 aiocb32_store_aiocb(struct aiocb **ujobp, struct aiocb *ujob)
2709 {
2710 
2711 	return (suword32(ujobp, (long)ujob));
2712 }
2713 
2714 static struct aiocb_ops aiocb32_ops = {
2715 	.copyin = aiocb32_copyin,
2716 	.fetch_status = aiocb32_fetch_status,
2717 	.fetch_error = aiocb32_fetch_error,
2718 	.store_status = aiocb32_store_status,
2719 	.store_error = aiocb32_store_error,
2720 	.store_kernelinfo = aiocb32_store_kernelinfo,
2721 	.store_aiocb = aiocb32_store_aiocb,
2722 };
2723 
2724 static struct aiocb_ops aiocb32_ops_osigevent = {
2725 	.copyin = aiocb32_copyin_old_sigevent,
2726 	.fetch_status = aiocb32_fetch_status,
2727 	.fetch_error = aiocb32_fetch_error,
2728 	.store_status = aiocb32_store_status,
2729 	.store_error = aiocb32_store_error,
2730 	.store_kernelinfo = aiocb32_store_kernelinfo,
2731 	.store_aiocb = aiocb32_store_aiocb,
2732 };
2733 
2734 int
2735 freebsd32_aio_return(struct thread *td, struct freebsd32_aio_return_args *uap)
2736 {
2737 
2738 	return (kern_aio_return(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2739 }
2740 
2741 int
2742 freebsd32_aio_suspend(struct thread *td, struct freebsd32_aio_suspend_args *uap)
2743 {
2744 	struct timespec32 ts32;
2745 	struct timespec ts, *tsp;
2746 	struct aiocb **ujoblist;
2747 	uint32_t *ujoblist32;
2748 	int error, i;
2749 
2750 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
2751 		return (EINVAL);
2752 
2753 	if (uap->timeout) {
2754 		/* Get timespec struct. */
2755 		if ((error = copyin(uap->timeout, &ts32, sizeof(ts32))) != 0)
2756 			return (error);
2757 		CP(ts32, ts, tv_sec);
2758 		CP(ts32, ts, tv_nsec);
2759 		tsp = &ts;
2760 	} else
2761 		tsp = NULL;
2762 
2763 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
2764 	ujoblist32 = (uint32_t *)ujoblist;
2765 	error = copyin(uap->aiocbp, ujoblist32, uap->nent *
2766 	    sizeof(ujoblist32[0]));
2767 	if (error == 0) {
2768 		for (i = uap->nent; i > 0; i--)
2769 			ujoblist[i] = PTRIN(ujoblist32[i]);
2770 
2771 		error = kern_aio_suspend(td, uap->nent, ujoblist, tsp);
2772 	}
2773 	uma_zfree(aiol_zone, ujoblist);
2774 	return (error);
2775 }
2776 
2777 int
2778 freebsd32_aio_cancel(struct thread *td, struct freebsd32_aio_cancel_args *uap)
2779 {
2780 
2781 	return (aio_cancel(td, (struct aio_cancel_args *)uap));
2782 }
2783 
2784 int
2785 freebsd32_aio_error(struct thread *td, struct freebsd32_aio_error_args *uap)
2786 {
2787 
2788 	return (kern_aio_error(td, (struct aiocb *)uap->aiocbp, &aiocb32_ops));
2789 }
2790 
2791 int
2792 freebsd32_oaio_read(struct thread *td, struct freebsd32_oaio_read_args *uap)
2793 {
2794 
2795 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2796 	    &aiocb32_ops_osigevent));
2797 }
2798 
2799 int
2800 freebsd32_aio_read(struct thread *td, struct freebsd32_aio_read_args *uap)
2801 {
2802 
2803 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ,
2804 	    &aiocb32_ops));
2805 }
2806 
2807 int
2808 freebsd32_oaio_write(struct thread *td, struct freebsd32_oaio_write_args *uap)
2809 {
2810 
2811 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2812 	    &aiocb32_ops_osigevent));
2813 }
2814 
2815 int
2816 freebsd32_aio_write(struct thread *td, struct freebsd32_aio_write_args *uap)
2817 {
2818 
2819 	return (aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE,
2820 	    &aiocb32_ops));
2821 }
2822 
2823 int
2824 freebsd32_aio_waitcomplete(struct thread *td,
2825     struct freebsd32_aio_waitcomplete_args *uap)
2826 {
2827 	struct timespec32 ts32;
2828 	struct timespec ts, *tsp;
2829 	int error;
2830 
2831 	if (uap->timeout) {
2832 		/* Get timespec struct. */
2833 		error = copyin(uap->timeout, &ts32, sizeof(ts32));
2834 		if (error)
2835 			return (error);
2836 		CP(ts32, ts, tv_sec);
2837 		CP(ts32, ts, tv_nsec);
2838 		tsp = &ts;
2839 	} else
2840 		tsp = NULL;
2841 
2842 	return (kern_aio_waitcomplete(td, (struct aiocb **)uap->aiocbp, tsp,
2843 	    &aiocb32_ops));
2844 }
2845 
2846 int
2847 freebsd32_aio_fsync(struct thread *td, struct freebsd32_aio_fsync_args *uap)
2848 {
2849 
2850 	return (kern_aio_fsync(td, uap->op, (struct aiocb *)uap->aiocbp,
2851 	    &aiocb32_ops));
2852 }
2853 
2854 int
2855 freebsd32_olio_listio(struct thread *td, struct freebsd32_olio_listio_args *uap)
2856 {
2857 	struct aiocb **acb_list;
2858 	struct sigevent *sigp, sig;
2859 	struct osigevent32 osig;
2860 	uint32_t *acb_list32;
2861 	int error, i, nent;
2862 
2863 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2864 		return (EINVAL);
2865 
2866 	nent = uap->nent;
2867 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2868 		return (EINVAL);
2869 
2870 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2871 		error = copyin(uap->sig, &osig, sizeof(osig));
2872 		if (error)
2873 			return (error);
2874 		error = convert_old_sigevent32(&osig, &sig);
2875 		if (error)
2876 			return (error);
2877 		sigp = &sig;
2878 	} else
2879 		sigp = NULL;
2880 
2881 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2882 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2883 	if (error) {
2884 		free(acb_list32, M_LIO);
2885 		return (error);
2886 	}
2887 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2888 	for (i = 0; i < nent; i++)
2889 		acb_list[i] = PTRIN(acb_list32[i]);
2890 	free(acb_list32, M_LIO);
2891 
2892 	error = kern_lio_listio(td, uap->mode,
2893 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2894 	    &aiocb32_ops_osigevent);
2895 	free(acb_list, M_LIO);
2896 	return (error);
2897 }
2898 
2899 int
2900 freebsd32_lio_listio(struct thread *td, struct freebsd32_lio_listio_args *uap)
2901 {
2902 	struct aiocb **acb_list;
2903 	struct sigevent *sigp, sig;
2904 	struct sigevent32 sig32;
2905 	uint32_t *acb_list32;
2906 	int error, i, nent;
2907 
2908 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
2909 		return (EINVAL);
2910 
2911 	nent = uap->nent;
2912 	if (nent < 0 || nent > AIO_LISTIO_MAX)
2913 		return (EINVAL);
2914 
2915 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
2916 		error = copyin(uap->sig, &sig32, sizeof(sig32));
2917 		if (error)
2918 			return (error);
2919 		error = convert_sigevent32(&sig32, &sig);
2920 		if (error)
2921 			return (error);
2922 		sigp = &sig;
2923 	} else
2924 		sigp = NULL;
2925 
2926 	acb_list32 = malloc(sizeof(uint32_t) * nent, M_LIO, M_WAITOK);
2927 	error = copyin(uap->acb_list, acb_list32, nent * sizeof(uint32_t));
2928 	if (error) {
2929 		free(acb_list32, M_LIO);
2930 		return (error);
2931 	}
2932 	acb_list = malloc(sizeof(struct aiocb *) * nent, M_LIO, M_WAITOK);
2933 	for (i = 0; i < nent; i++)
2934 		acb_list[i] = PTRIN(acb_list32[i]);
2935 	free(acb_list32, M_LIO);
2936 
2937 	error = kern_lio_listio(td, uap->mode,
2938 	    (struct aiocb * const *)uap->acb_list, acb_list, nent, sigp,
2939 	    &aiocb32_ops);
2940 	free(acb_list, M_LIO);
2941 	return (error);
2942 }
2943 
2944 SYSCALL32_MODULE_HELPER(freebsd32_aio_return);
2945 SYSCALL32_MODULE_HELPER(freebsd32_aio_suspend);
2946 SYSCALL32_MODULE_HELPER(freebsd32_aio_cancel);
2947 SYSCALL32_MODULE_HELPER(freebsd32_aio_error);
2948 SYSCALL32_MODULE_HELPER(freebsd32_aio_fsync);
2949 SYSCALL32_MODULE_HELPER(freebsd32_aio_read);
2950 SYSCALL32_MODULE_HELPER(freebsd32_aio_write);
2951 SYSCALL32_MODULE_HELPER(freebsd32_aio_waitcomplete);
2952 SYSCALL32_MODULE_HELPER(freebsd32_lio_listio);
2953 SYSCALL32_MODULE_HELPER(freebsd32_oaio_read);
2954 SYSCALL32_MODULE_HELPER(freebsd32_oaio_write);
2955 SYSCALL32_MODULE_HELPER(freebsd32_olio_listio);
2956 #endif
2957