xref: /freebsd/sys/kern/vfs_aio.c (revision 282a3889ebf826db9839be296ff1dd903f6d6d6e)
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 <sys/param.h>
25 #include <sys/systm.h>
26 #include <sys/malloc.h>
27 #include <sys/bio.h>
28 #include <sys/buf.h>
29 #include <sys/eventhandler.h>
30 #include <sys/sysproto.h>
31 #include <sys/filedesc.h>
32 #include <sys/kernel.h>
33 #include <sys/module.h>
34 #include <sys/kthread.h>
35 #include <sys/fcntl.h>
36 #include <sys/file.h>
37 #include <sys/limits.h>
38 #include <sys/lock.h>
39 #include <sys/mutex.h>
40 #include <sys/unistd.h>
41 #include <sys/posix4.h>
42 #include <sys/proc.h>
43 #include <sys/resourcevar.h>
44 #include <sys/signalvar.h>
45 #include <sys/protosw.h>
46 #include <sys/sema.h>
47 #include <sys/socket.h>
48 #include <sys/socketvar.h>
49 #include <sys/syscall.h>
50 #include <sys/sysent.h>
51 #include <sys/sysctl.h>
52 #include <sys/sx.h>
53 #include <sys/taskqueue.h>
54 #include <sys/vnode.h>
55 #include <sys/conf.h>
56 #include <sys/event.h>
57 #include <sys/mount.h>
58 
59 #include <machine/atomic.h>
60 
61 #include <vm/vm.h>
62 #include <vm/vm_extern.h>
63 #include <vm/pmap.h>
64 #include <vm/vm_map.h>
65 #include <vm/vm_object.h>
66 #include <vm/uma.h>
67 #include <sys/aio.h>
68 
69 #include "opt_vfs_aio.h"
70 
71 /*
72  * Counter for allocating reference ids to new jobs.  Wrapped to 1 on
73  * overflow. (XXX will be removed soon.)
74  */
75 static u_long jobrefid;
76 
77 /*
78  * Counter for aio_fsync.
79  */
80 static uint64_t jobseqno;
81 
82 #define JOBST_NULL		0
83 #define JOBST_JOBQSOCK		1
84 #define JOBST_JOBQGLOBAL	2
85 #define JOBST_JOBRUNNING	3
86 #define JOBST_JOBFINISHED	4
87 #define JOBST_JOBQBUF		5
88 #define JOBST_JOBQSYNC		6
89 
90 #ifndef MAX_AIO_PER_PROC
91 #define MAX_AIO_PER_PROC	32
92 #endif
93 
94 #ifndef MAX_AIO_QUEUE_PER_PROC
95 #define MAX_AIO_QUEUE_PER_PROC	256 /* Bigger than AIO_LISTIO_MAX */
96 #endif
97 
98 #ifndef MAX_AIO_PROCS
99 #define MAX_AIO_PROCS		32
100 #endif
101 
102 #ifndef MAX_AIO_QUEUE
103 #define	MAX_AIO_QUEUE		1024 /* Bigger than AIO_LISTIO_MAX */
104 #endif
105 
106 #ifndef TARGET_AIO_PROCS
107 #define TARGET_AIO_PROCS	4
108 #endif
109 
110 #ifndef MAX_BUF_AIO
111 #define MAX_BUF_AIO		16
112 #endif
113 
114 #ifndef AIOD_TIMEOUT_DEFAULT
115 #define	AIOD_TIMEOUT_DEFAULT	(10 * hz)
116 #endif
117 
118 #ifndef AIOD_LIFETIME_DEFAULT
119 #define AIOD_LIFETIME_DEFAULT	(30 * hz)
120 #endif
121 
122 static SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
123 
124 static int max_aio_procs = MAX_AIO_PROCS;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
126 	CTLFLAG_RW, &max_aio_procs, 0,
127 	"Maximum number of kernel threads to use for handling async IO ");
128 
129 static int num_aio_procs = 0;
130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
131 	CTLFLAG_RD, &num_aio_procs, 0,
132 	"Number of presently active kernel threads for async IO");
133 
134 /*
135  * The code will adjust the actual number of AIO processes towards this
136  * number when it gets a chance.
137  */
138 static int target_aio_procs = TARGET_AIO_PROCS;
139 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
140 	0, "Preferred number of ready kernel threads for async IO");
141 
142 static int max_queue_count = MAX_AIO_QUEUE;
143 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
144     "Maximum number of aio requests to queue, globally");
145 
146 static int num_queue_count = 0;
147 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
148     "Number of queued aio requests");
149 
150 static int num_buf_aio = 0;
151 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
152     "Number of aio requests presently handled by the buf subsystem");
153 
154 /* Number of async I/O thread in the process of being started */
155 /* XXX This should be local to aio_aqueue() */
156 static int num_aio_resv_start = 0;
157 
158 static int aiod_timeout;
159 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
160     "Timeout value for synchronous aio operations");
161 
162 static int aiod_lifetime;
163 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
164     "Maximum lifetime for idle aiod");
165 
166 static int unloadable = 0;
167 SYSCTL_INT(_vfs_aio, OID_AUTO, unloadable, CTLFLAG_RW, &unloadable, 0,
168     "Allow unload of aio (not recommended)");
169 
170 
171 static int max_aio_per_proc = MAX_AIO_PER_PROC;
172 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
173     0, "Maximum active aio requests per process (stored in the process)");
174 
175 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
176 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
177     &max_aio_queue_per_proc, 0,
178     "Maximum queued aio requests per process (stored in the process)");
179 
180 static int max_buf_aio = MAX_BUF_AIO;
181 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
182     "Maximum buf aio requests per process (stored in the process)");
183 
184 typedef struct oaiocb {
185 	int	aio_fildes;		/* File descriptor */
186 	off_t	aio_offset;		/* File offset for I/O */
187 	volatile void *aio_buf;         /* I/O buffer in process space */
188 	size_t	aio_nbytes;		/* Number of bytes for I/O */
189 	struct	osigevent aio_sigevent;	/* Signal to deliver */
190 	int	aio_lio_opcode;		/* LIO opcode */
191 	int	aio_reqprio;		/* Request priority -- ignored */
192 	struct	__aiocb_private	_aiocb_private;
193 } oaiocb_t;
194 
195 /*
196  * Below is a key of locks used to protect each member of struct aiocblist
197  * aioliojob and kaioinfo and any backends.
198  *
199  * * - need not protected
200  * a - locked by kaioinfo lock
201  * b - locked by backend lock, the backend lock can be null in some cases,
202  *     for example, BIO belongs to this type, in this case, proc lock is
203  *     reused.
204  * c - locked by aio_job_mtx, the lock for the generic file I/O backend.
205  */
206 
207 /*
208  * Current, there is only two backends: BIO and generic file I/O.
209  * socket I/O is served by generic file I/O, this is not a good idea, since
210  * disk file I/O and any other types without O_NONBLOCK flag can block daemon
211  * threads, if there is no thread to serve socket I/O, the socket I/O will be
212  * delayed too long or starved, we should create some threads dedicated to
213  * sockets to do non-blocking I/O, same for pipe and fifo, for these I/O
214  * systems we really need non-blocking interface, fiddling O_NONBLOCK in file
215  * structure is not safe because there is race between userland and aio
216  * daemons.
217  */
218 
219 struct aiocblist {
220 	TAILQ_ENTRY(aiocblist) list;	/* (b) internal list of for backend */
221 	TAILQ_ENTRY(aiocblist) plist;	/* (a) list of jobs for each backend */
222 	TAILQ_ENTRY(aiocblist) allist;  /* (a) list of all jobs in proc */
223 	int	jobflags;		/* (a) job flags */
224 	int	jobstate;		/* (b) job state */
225 	int	inputcharge;		/* (*) input blockes */
226 	int	outputcharge;		/* (*) output blockes */
227 	struct	buf *bp;		/* (*) private to BIO backend,
228 				  	 * buffer pointer
229 					 */
230 	struct	proc *userproc;		/* (*) user process */
231 	struct  ucred *cred;		/* (*) active credential when created */
232 	struct	file *fd_file;		/* (*) pointer to file structure */
233 	struct	aioliojob *lio;		/* (*) optional lio job */
234 	struct	aiocb *uuaiocb;		/* (*) pointer in userspace of aiocb */
235 	struct	knlist klist;		/* (a) list of knotes */
236 	struct	aiocb uaiocb;		/* (*) kernel I/O control block */
237 	ksiginfo_t ksi;			/* (a) realtime signal info */
238 	struct	task biotask;		/* (*) private to BIO backend */
239 	uint64_t seqno;			/* (*) job number */
240 	int	pending;		/* (a) number of pending I/O, aio_fsync only */
241 };
242 
243 /* jobflags */
244 #define AIOCBLIST_DONE		0x01
245 #define AIOCBLIST_BUFDONE	0x02
246 #define AIOCBLIST_RUNDOWN	0x04
247 #define AIOCBLIST_CHECKSYNC	0x08
248 
249 /*
250  * AIO process info
251  */
252 #define AIOP_FREE	0x1			/* proc on free queue */
253 
254 struct aiothreadlist {
255 	int aiothreadflags;			/* (c) AIO proc flags */
256 	TAILQ_ENTRY(aiothreadlist) list;	/* (c) list of processes */
257 	struct thread *aiothread;		/* (*) the AIO thread */
258 };
259 
260 /*
261  * data-structure for lio signal management
262  */
263 struct aioliojob {
264 	int	lioj_flags;			/* (a) listio flags */
265 	int	lioj_count;			/* (a) listio flags */
266 	int	lioj_finished_count;		/* (a) listio flags */
267 	struct	sigevent lioj_signal;		/* (a) signal on all I/O done */
268 	TAILQ_ENTRY(aioliojob) lioj_list;	/* (a) lio list */
269 	struct  knlist klist;			/* (a) list of knotes */
270 	ksiginfo_t lioj_ksi;			/* (a) Realtime signal info */
271 };
272 
273 #define	LIOJ_SIGNAL		0x1	/* signal on all done (lio) */
274 #define	LIOJ_SIGNAL_POSTED	0x2	/* signal has been posted */
275 #define LIOJ_KEVENT_POSTED	0x4	/* kevent triggered */
276 
277 /*
278  * per process aio data structure
279  */
280 struct kaioinfo {
281 	struct mtx	kaio_mtx;	/* the lock to protect this struct */
282 	int	kaio_flags;		/* (a) per process kaio flags */
283 	int	kaio_maxactive_count;	/* (*) maximum number of AIOs */
284 	int	kaio_active_count;	/* (c) number of currently used AIOs */
285 	int	kaio_qallowed_count;	/* (*) maxiumu size of AIO queue */
286 	int	kaio_count;		/* (a) size of AIO queue */
287 	int	kaio_ballowed_count;	/* (*) maximum number of buffers */
288 	int	kaio_buffer_count;	/* (a) number of physio buffers */
289 	TAILQ_HEAD(,aiocblist) kaio_all;	/* (a) all AIOs in the process */
290 	TAILQ_HEAD(,aiocblist) kaio_done;	/* (a) done queue for process */
291 	TAILQ_HEAD(,aioliojob) kaio_liojoblist; /* (a) list of lio jobs */
292 	TAILQ_HEAD(,aiocblist) kaio_jobqueue;	/* (a) job queue for process */
293 	TAILQ_HEAD(,aiocblist) kaio_bufqueue;	/* (a) buffer job queue for process */
294 	TAILQ_HEAD(,aiocblist) kaio_sockqueue;  /* (a) queue for aios waiting on sockets,
295 						 *  NOT USED YET.
296 						 */
297 	TAILQ_HEAD(,aiocblist) kaio_syncqueue;	/* (a) queue for aio_fsync */
298 	struct	task	kaio_task;	/* (*) task to kick aio threads */
299 };
300 
301 #define AIO_LOCK(ki)		mtx_lock(&(ki)->kaio_mtx)
302 #define AIO_UNLOCK(ki)		mtx_unlock(&(ki)->kaio_mtx)
303 #define AIO_LOCK_ASSERT(ki, f)	mtx_assert(&(ki)->kaio_mtx, (f))
304 #define AIO_MTX(ki)		(&(ki)->kaio_mtx)
305 
306 #define KAIO_RUNDOWN	0x1	/* process is being run down */
307 #define KAIO_WAKEUP	0x2	/* wakeup process when there is a significant event */
308 
309 static TAILQ_HEAD(,aiothreadlist) aio_freeproc;		/* (c) Idle daemons */
310 static struct sema aio_newproc_sem;
311 static struct mtx aio_job_mtx;
312 static struct mtx aio_sock_mtx;
313 static TAILQ_HEAD(,aiocblist) aio_jobs;			/* (c) Async job list */
314 static struct unrhdr *aiod_unr;
315 
316 void		aio_init_aioinfo(struct proc *p);
317 static void	aio_onceonly(void);
318 static int	aio_free_entry(struct aiocblist *aiocbe);
319 static void	aio_process(struct aiocblist *aiocbe);
320 static int	aio_newproc(int *);
321 int		aio_aqueue(struct thread *td, struct aiocb *job,
322 			struct aioliojob *lio, int type, int osigev);
323 static void	aio_physwakeup(struct buf *bp);
324 static void	aio_proc_rundown(void *arg, struct proc *p);
325 static void	aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp);
326 static int	aio_qphysio(struct proc *p, struct aiocblist *iocb);
327 static void	biohelper(void *, int);
328 static void	aio_daemon(void *param);
329 static void	aio_swake_cb(struct socket *, struct sockbuf *);
330 static int	aio_unload(void);
331 static void	aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type);
332 #define DONE_BUF	1
333 #define DONE_QUEUE	2
334 static int	do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev);
335 static int	aio_kick(struct proc *userp);
336 static void	aio_kick_nowait(struct proc *userp);
337 static void	aio_kick_helper(void *context, int pending);
338 static int	filt_aioattach(struct knote *kn);
339 static void	filt_aiodetach(struct knote *kn);
340 static int	filt_aio(struct knote *kn, long hint);
341 static int	filt_lioattach(struct knote *kn);
342 static void	filt_liodetach(struct knote *kn);
343 static int	filt_lio(struct knote *kn, long hint);
344 
345 /*
346  * Zones for:
347  * 	kaio	Per process async io info
348  *	aiop	async io thread data
349  *	aiocb	async io jobs
350  *	aiol	list io job pointer - internal to aio_suspend XXX
351  *	aiolio	list io jobs
352  */
353 static uma_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
354 
355 /* kqueue filters for aio */
356 static struct filterops aio_filtops =
357 	{ 0, filt_aioattach, filt_aiodetach, filt_aio };
358 static struct filterops lio_filtops =
359 	{ 0, filt_lioattach, filt_liodetach, filt_lio };
360 
361 static eventhandler_tag exit_tag, exec_tag;
362 
363 TASKQUEUE_DEFINE_THREAD(aiod_bio);
364 
365 /*
366  * Main operations function for use as a kernel module.
367  */
368 static int
369 aio_modload(struct module *module, int cmd, void *arg)
370 {
371 	int error = 0;
372 
373 	switch (cmd) {
374 	case MOD_LOAD:
375 		aio_onceonly();
376 		break;
377 	case MOD_UNLOAD:
378 		error = aio_unload();
379 		break;
380 	case MOD_SHUTDOWN:
381 		break;
382 	default:
383 		error = EINVAL;
384 		break;
385 	}
386 	return (error);
387 }
388 
389 static moduledata_t aio_mod = {
390 	"aio",
391 	&aio_modload,
392 	NULL
393 };
394 
395 SYSCALL_MODULE_HELPER(aio_cancel);
396 SYSCALL_MODULE_HELPER(aio_error);
397 SYSCALL_MODULE_HELPER(aio_fsync);
398 SYSCALL_MODULE_HELPER(aio_read);
399 SYSCALL_MODULE_HELPER(aio_return);
400 SYSCALL_MODULE_HELPER(aio_suspend);
401 SYSCALL_MODULE_HELPER(aio_waitcomplete);
402 SYSCALL_MODULE_HELPER(aio_write);
403 SYSCALL_MODULE_HELPER(lio_listio);
404 SYSCALL_MODULE_HELPER(oaio_read);
405 SYSCALL_MODULE_HELPER(oaio_write);
406 SYSCALL_MODULE_HELPER(olio_listio);
407 
408 DECLARE_MODULE(aio, aio_mod,
409 	SI_SUB_VFS, SI_ORDER_ANY);
410 MODULE_VERSION(aio, 1);
411 
412 /*
413  * Startup initialization
414  */
415 static void
416 aio_onceonly(void)
417 {
418 
419 	/* XXX: should probably just use so->callback */
420 	aio_swake = &aio_swake_cb;
421 	exit_tag = EVENTHANDLER_REGISTER(process_exit, aio_proc_rundown, NULL,
422 	    EVENTHANDLER_PRI_ANY);
423 	exec_tag = EVENTHANDLER_REGISTER(process_exec, aio_proc_rundown_exec, NULL,
424 	    EVENTHANDLER_PRI_ANY);
425 	kqueue_add_filteropts(EVFILT_AIO, &aio_filtops);
426 	kqueue_add_filteropts(EVFILT_LIO, &lio_filtops);
427 	TAILQ_INIT(&aio_freeproc);
428 	sema_init(&aio_newproc_sem, 0, "aio_new_proc");
429 	mtx_init(&aio_job_mtx, "aio_job", NULL, MTX_DEF);
430 	mtx_init(&aio_sock_mtx, "aio_sock", NULL, MTX_DEF);
431 	TAILQ_INIT(&aio_jobs);
432 	aiod_unr = new_unrhdr(1, INT_MAX, NULL);
433 	kaio_zone = uma_zcreate("AIO", sizeof(struct kaioinfo), NULL, NULL,
434 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
435 	aiop_zone = uma_zcreate("AIOP", sizeof(struct aiothreadlist), NULL,
436 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
437 	aiocb_zone = uma_zcreate("AIOCB", sizeof(struct aiocblist), NULL, NULL,
438 	    NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
439 	aiol_zone = uma_zcreate("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t) , NULL,
440 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
441 	aiolio_zone = uma_zcreate("AIOLIO", sizeof(struct aioliojob), NULL,
442 	    NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE);
443 	aiod_timeout = AIOD_TIMEOUT_DEFAULT;
444 	aiod_lifetime = AIOD_LIFETIME_DEFAULT;
445 	jobrefid = 1;
446 	async_io_version = _POSIX_VERSION;
447 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, AIO_LISTIO_MAX);
448 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, MAX_AIO_QUEUE);
449 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, 0);
450 }
451 
452 /*
453  * Callback for unload of AIO when used as a module.
454  */
455 static int
456 aio_unload(void)
457 {
458 	int error;
459 
460 	/*
461 	 * XXX: no unloads by default, it's too dangerous.
462 	 * perhaps we could do it if locked out callers and then
463 	 * did an aio_proc_rundown() on each process.
464 	 *
465 	 * jhb: aio_proc_rundown() needs to run on curproc though,
466 	 * so I don't think that would fly.
467 	 */
468 	if (!unloadable)
469 		return (EOPNOTSUPP);
470 
471 	error = kqueue_del_filteropts(EVFILT_AIO);
472 	if (error)
473 		return error;
474 	error = kqueue_del_filteropts(EVFILT_LIO);
475 	if (error)
476 		return error;
477 	async_io_version = 0;
478 	aio_swake = NULL;
479 	taskqueue_free(taskqueue_aiod_bio);
480 	delete_unrhdr(aiod_unr);
481 	uma_zdestroy(kaio_zone);
482 	uma_zdestroy(aiop_zone);
483 	uma_zdestroy(aiocb_zone);
484 	uma_zdestroy(aiol_zone);
485 	uma_zdestroy(aiolio_zone);
486 	EVENTHANDLER_DEREGISTER(process_exit, exit_tag);
487 	EVENTHANDLER_DEREGISTER(process_exec, exec_tag);
488 	mtx_destroy(&aio_job_mtx);
489 	mtx_destroy(&aio_sock_mtx);
490 	sema_destroy(&aio_newproc_sem);
491 	p31b_setcfg(CTL_P1003_1B_AIO_LISTIO_MAX, -1);
492 	p31b_setcfg(CTL_P1003_1B_AIO_MAX, -1);
493 	p31b_setcfg(CTL_P1003_1B_AIO_PRIO_DELTA_MAX, -1);
494 	return (0);
495 }
496 
497 /*
498  * Init the per-process aioinfo structure.  The aioinfo limits are set
499  * per-process for user limit (resource) management.
500  */
501 void
502 aio_init_aioinfo(struct proc *p)
503 {
504 	struct kaioinfo *ki;
505 
506 	ki = uma_zalloc(kaio_zone, M_WAITOK);
507 	mtx_init(&ki->kaio_mtx, "aiomtx", NULL, MTX_DEF);
508 	ki->kaio_flags = 0;
509 	ki->kaio_maxactive_count = max_aio_per_proc;
510 	ki->kaio_active_count = 0;
511 	ki->kaio_qallowed_count = max_aio_queue_per_proc;
512 	ki->kaio_count = 0;
513 	ki->kaio_ballowed_count = max_buf_aio;
514 	ki->kaio_buffer_count = 0;
515 	TAILQ_INIT(&ki->kaio_all);
516 	TAILQ_INIT(&ki->kaio_done);
517 	TAILQ_INIT(&ki->kaio_jobqueue);
518 	TAILQ_INIT(&ki->kaio_bufqueue);
519 	TAILQ_INIT(&ki->kaio_liojoblist);
520 	TAILQ_INIT(&ki->kaio_sockqueue);
521 	TAILQ_INIT(&ki->kaio_syncqueue);
522 	TASK_INIT(&ki->kaio_task, 0, aio_kick_helper, p);
523 	PROC_LOCK(p);
524 	if (p->p_aioinfo == NULL) {
525 		p->p_aioinfo = ki;
526 		PROC_UNLOCK(p);
527 	} else {
528 		PROC_UNLOCK(p);
529 		mtx_destroy(&ki->kaio_mtx);
530 		uma_zfree(kaio_zone, ki);
531 	}
532 
533 	while (num_aio_procs < target_aio_procs)
534 		aio_newproc(NULL);
535 }
536 
537 static int
538 aio_sendsig(struct proc *p, struct sigevent *sigev, ksiginfo_t *ksi)
539 {
540 	int ret = 0;
541 
542 	PROC_LOCK(p);
543 	if (!KSI_ONQ(ksi)) {
544 		ksi->ksi_code = SI_ASYNCIO;
545 		ksi->ksi_flags |= KSI_EXT | KSI_INS;
546 		ret = psignal_event(p, sigev, ksi);
547 	}
548 	PROC_UNLOCK(p);
549 	return (ret);
550 }
551 
552 /*
553  * Free a job entry.  Wait for completion if it is currently active, but don't
554  * delay forever.  If we delay, we return a flag that says that we have to
555  * restart the queue scan.
556  */
557 static int
558 aio_free_entry(struct aiocblist *aiocbe)
559 {
560 	struct kaioinfo *ki;
561 	struct aioliojob *lj;
562 	struct proc *p;
563 
564 	p = aiocbe->userproc;
565 	MPASS(curproc == p);
566 	ki = p->p_aioinfo;
567 	MPASS(ki != NULL);
568 
569 	AIO_LOCK_ASSERT(ki, MA_OWNED);
570 	MPASS(aiocbe->jobstate == JOBST_JOBFINISHED);
571 
572 	atomic_subtract_int(&num_queue_count, 1);
573 
574 	ki->kaio_count--;
575 	MPASS(ki->kaio_count >= 0);
576 
577 	TAILQ_REMOVE(&ki->kaio_done, aiocbe, plist);
578 	TAILQ_REMOVE(&ki->kaio_all, aiocbe, allist);
579 
580 	lj = aiocbe->lio;
581 	if (lj) {
582 		lj->lioj_count--;
583 		lj->lioj_finished_count--;
584 
585 		if (lj->lioj_count == 0) {
586 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
587 			/* lio is going away, we need to destroy any knotes */
588 			knlist_delete(&lj->klist, curthread, 1);
589 			PROC_LOCK(p);
590 			sigqueue_take(&lj->lioj_ksi);
591 			PROC_UNLOCK(p);
592 			uma_zfree(aiolio_zone, lj);
593 		}
594 	}
595 
596 	/* aiocbe is going away, we need to destroy any knotes */
597 	knlist_delete(&aiocbe->klist, curthread, 1);
598 	PROC_LOCK(p);
599 	sigqueue_take(&aiocbe->ksi);
600 	PROC_UNLOCK(p);
601 
602 	MPASS(aiocbe->bp == NULL);
603 	aiocbe->jobstate = JOBST_NULL;
604 	AIO_UNLOCK(ki);
605 
606 	/*
607 	 * The thread argument here is used to find the owning process
608 	 * and is also passed to fo_close() which may pass it to various
609 	 * places such as devsw close() routines.  Because of that, we
610 	 * need a thread pointer from the process owning the job that is
611 	 * persistent and won't disappear out from under us or move to
612 	 * another process.
613 	 *
614 	 * Currently, all the callers of this function call it to remove
615 	 * an aiocblist from the current process' job list either via a
616 	 * syscall or due to the current process calling exit() or
617 	 * execve().  Thus, we know that p == curproc.  We also know that
618 	 * curthread can't exit since we are curthread.
619 	 *
620 	 * Therefore, we use curthread as the thread to pass to
621 	 * knlist_delete().  This does mean that it is possible for the
622 	 * thread pointer at close time to differ from the thread pointer
623 	 * at open time, but this is already true of file descriptors in
624 	 * a multithreaded process.
625 	 */
626 	fdrop(aiocbe->fd_file, curthread);
627 	crfree(aiocbe->cred);
628 	uma_zfree(aiocb_zone, aiocbe);
629 	AIO_LOCK(ki);
630 
631 	return (0);
632 }
633 
634 static void
635 aio_proc_rundown_exec(void *arg, struct proc *p, struct image_params *imgp __unused)
636 {
637    	aio_proc_rundown(arg, p);
638 }
639 
640 /*
641  * Rundown the jobs for a given process.
642  */
643 static void
644 aio_proc_rundown(void *arg, struct proc *p)
645 {
646 	struct kaioinfo *ki;
647 	struct aioliojob *lj;
648 	struct aiocblist *cbe, *cbn;
649 	struct file *fp;
650 	struct socket *so;
651 	int remove;
652 
653 	KASSERT(curthread->td_proc == p,
654 	    ("%s: called on non-curproc", __func__));
655 	ki = p->p_aioinfo;
656 	if (ki == NULL)
657 		return;
658 
659 	AIO_LOCK(ki);
660 	ki->kaio_flags |= KAIO_RUNDOWN;
661 
662 restart:
663 
664 	/*
665 	 * Try to cancel all pending requests. This code simulates
666 	 * aio_cancel on all pending I/O requests.
667 	 */
668 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
669 		remove = 0;
670 		mtx_lock(&aio_job_mtx);
671 		if (cbe->jobstate == JOBST_JOBQGLOBAL) {
672 			TAILQ_REMOVE(&aio_jobs, cbe, list);
673 			remove = 1;
674 		} else if (cbe->jobstate == JOBST_JOBQSOCK) {
675 			fp = cbe->fd_file;
676 			MPASS(fp->f_type == DTYPE_SOCKET);
677 			so = fp->f_data;
678 			TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
679 			remove = 1;
680 		} else if (cbe->jobstate == JOBST_JOBQSYNC) {
681 			TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
682 			remove = 1;
683 		}
684 		mtx_unlock(&aio_job_mtx);
685 
686 		if (remove) {
687 			cbe->jobstate = JOBST_JOBFINISHED;
688 			cbe->uaiocb._aiocb_private.status = -1;
689 			cbe->uaiocb._aiocb_private.error = ECANCELED;
690 			TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
691 			aio_bio_done_notify(p, cbe, DONE_QUEUE);
692 		}
693 	}
694 
695 	/* Wait for all running I/O to be finished */
696 	if (TAILQ_FIRST(&ki->kaio_bufqueue) ||
697 	    TAILQ_FIRST(&ki->kaio_jobqueue)) {
698 		ki->kaio_flags |= KAIO_WAKEUP;
699 		msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO, "aioprn", hz);
700 		goto restart;
701 	}
702 
703 	/* Free all completed I/O requests. */
704 	while ((cbe = TAILQ_FIRST(&ki->kaio_done)) != NULL)
705 		aio_free_entry(cbe);
706 
707 	while ((lj = TAILQ_FIRST(&ki->kaio_liojoblist)) != NULL) {
708 		if (lj->lioj_count == 0) {
709 			TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
710 			knlist_delete(&lj->klist, curthread, 1);
711 			PROC_LOCK(p);
712 			sigqueue_take(&lj->lioj_ksi);
713 			PROC_UNLOCK(p);
714 			uma_zfree(aiolio_zone, lj);
715 		} else {
716 			panic("LIO job not cleaned up: C:%d, FC:%d\n",
717 			    lj->lioj_count, lj->lioj_finished_count);
718 		}
719 	}
720 	AIO_UNLOCK(ki);
721 	taskqueue_drain(taskqueue_aiod_bio, &ki->kaio_task);
722 	uma_zfree(kaio_zone, ki);
723 	p->p_aioinfo = NULL;
724 }
725 
726 /*
727  * Select a job to run (called by an AIO daemon).
728  */
729 static struct aiocblist *
730 aio_selectjob(struct aiothreadlist *aiop)
731 {
732 	struct aiocblist *aiocbe;
733 	struct kaioinfo *ki;
734 	struct proc *userp;
735 
736 	mtx_assert(&aio_job_mtx, MA_OWNED);
737 	TAILQ_FOREACH(aiocbe, &aio_jobs, list) {
738 		userp = aiocbe->userproc;
739 		ki = userp->p_aioinfo;
740 
741 		if (ki->kaio_active_count < ki->kaio_maxactive_count) {
742 			TAILQ_REMOVE(&aio_jobs, aiocbe, list);
743 			/* Account for currently active jobs. */
744 			ki->kaio_active_count++;
745 			aiocbe->jobstate = JOBST_JOBRUNNING;
746 			break;
747 		}
748 	}
749 	return (aiocbe);
750 }
751 
752 /*
753  *  Move all data to a permanent storage device, this code
754  *  simulates fsync syscall.
755  */
756 static int
757 aio_fsync_vnode(struct thread *td, struct vnode *vp)
758 {
759 	struct mount *mp;
760 	int vfslocked;
761 	int error;
762 
763 	vfslocked = VFS_LOCK_GIANT(vp->v_mount);
764 	if ((error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0)
765 		goto drop;
766 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY, td);
767 	if (vp->v_object != NULL) {
768 		VM_OBJECT_LOCK(vp->v_object);
769 		vm_object_page_clean(vp->v_object, 0, 0, 0);
770 		VM_OBJECT_UNLOCK(vp->v_object);
771 	}
772 	error = VOP_FSYNC(vp, MNT_WAIT, td);
773 
774 	VOP_UNLOCK(vp, 0, td);
775 	vn_finished_write(mp);
776 drop:
777 	VFS_UNLOCK_GIANT(vfslocked);
778 	return (error);
779 }
780 
781 /*
782  * The AIO processing activity.  This is the code that does the I/O request for
783  * the non-physio version of the operations.  The normal vn operations are used,
784  * and this code should work in all instances for every type of file, including
785  * pipes, sockets, fifos, and regular files.
786  *
787  * XXX I don't think it works well for socket, pipe, and fifo.
788  */
789 static void
790 aio_process(struct aiocblist *aiocbe)
791 {
792 	struct ucred *td_savedcred;
793 	struct thread *td;
794 	struct aiocb *cb;
795 	struct file *fp;
796 	struct socket *so;
797 	struct uio auio;
798 	struct iovec aiov;
799 	int cnt;
800 	int error;
801 	int oublock_st, oublock_end;
802 	int inblock_st, inblock_end;
803 
804 	td = curthread;
805 	td_savedcred = td->td_ucred;
806 	td->td_ucred = aiocbe->cred;
807 	cb = &aiocbe->uaiocb;
808 	fp = aiocbe->fd_file;
809 
810 	if (cb->aio_lio_opcode == LIO_SYNC) {
811 		error = 0;
812 		cnt = 0;
813 		if (fp->f_vnode != NULL)
814 			error = aio_fsync_vnode(td, fp->f_vnode);
815 		cb->_aiocb_private.error = error;
816 		cb->_aiocb_private.status = 0;
817 		td->td_ucred = td_savedcred;
818 		return;
819 	}
820 
821 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
822 	aiov.iov_len = cb->aio_nbytes;
823 
824 	auio.uio_iov = &aiov;
825 	auio.uio_iovcnt = 1;
826 	auio.uio_offset = cb->aio_offset;
827 	auio.uio_resid = cb->aio_nbytes;
828 	cnt = cb->aio_nbytes;
829 	auio.uio_segflg = UIO_USERSPACE;
830 	auio.uio_td = td;
831 
832 	inblock_st = td->td_ru.ru_inblock;
833 	oublock_st = td->td_ru.ru_oublock;
834 	/*
835 	 * aio_aqueue() acquires a reference to the file that is
836 	 * released in aio_free_entry().
837 	 */
838 	if (cb->aio_lio_opcode == LIO_READ) {
839 		auio.uio_rw = UIO_READ;
840 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
841 	} else {
842 		if (fp->f_type == DTYPE_VNODE)
843 			bwillwrite();
844 		auio.uio_rw = UIO_WRITE;
845 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
846 	}
847 	inblock_end = td->td_ru.ru_inblock;
848 	oublock_end = td->td_ru.ru_oublock;
849 
850 	aiocbe->inputcharge = inblock_end - inblock_st;
851 	aiocbe->outputcharge = oublock_end - oublock_st;
852 
853 	if ((error) && (auio.uio_resid != cnt)) {
854 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
855 			error = 0;
856 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
857 			int sigpipe = 1;
858 			if (fp->f_type == DTYPE_SOCKET) {
859 				so = fp->f_data;
860 				if (so->so_options & SO_NOSIGPIPE)
861 					sigpipe = 0;
862 			}
863 			if (sigpipe) {
864 				PROC_LOCK(aiocbe->userproc);
865 				psignal(aiocbe->userproc, SIGPIPE);
866 				PROC_UNLOCK(aiocbe->userproc);
867 			}
868 		}
869 	}
870 
871 	cnt -= auio.uio_resid;
872 	cb->_aiocb_private.error = error;
873 	cb->_aiocb_private.status = cnt;
874 	td->td_ucred = td_savedcred;
875 }
876 
877 static void
878 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
879 {
880 	struct aioliojob *lj;
881 	struct kaioinfo *ki;
882 	struct aiocblist *scb, *scbn;
883 	int lj_done;
884 
885 	ki = userp->p_aioinfo;
886 	AIO_LOCK_ASSERT(ki, MA_OWNED);
887 	lj = aiocbe->lio;
888 	lj_done = 0;
889 	if (lj) {
890 		lj->lioj_finished_count++;
891 		if (lj->lioj_count == lj->lioj_finished_count)
892 			lj_done = 1;
893 	}
894 	if (type == DONE_QUEUE) {
895 		aiocbe->jobflags |= AIOCBLIST_DONE;
896 	} else {
897 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
898 	}
899 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
900 	aiocbe->jobstate = JOBST_JOBFINISHED;
901 
902 	if (ki->kaio_flags & KAIO_RUNDOWN)
903 		goto notification_done;
904 
905 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
906 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
907 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
908 
909 	KNOTE_LOCKED(&aiocbe->klist, 1);
910 
911 	if (lj_done) {
912 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
913 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
914 			KNOTE_LOCKED(&lj->klist, 1);
915 		}
916 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
917 		    == LIOJ_SIGNAL
918 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
919 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
920 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
921 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
922 		}
923 	}
924 
925 notification_done:
926 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
927 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
928 			if (aiocbe->fd_file == scb->fd_file &&
929 			    aiocbe->seqno < scb->seqno) {
930 				if (--scb->pending == 0) {
931 					mtx_lock(&aio_job_mtx);
932 					scb->jobstate = JOBST_JOBQGLOBAL;
933 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
934 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
935 					aio_kick_nowait(userp);
936 					mtx_unlock(&aio_job_mtx);
937 				}
938 			}
939 		}
940 	}
941 	if (ki->kaio_flags & KAIO_WAKEUP) {
942 		ki->kaio_flags &= ~KAIO_WAKEUP;
943 		wakeup(&userp->p_aioinfo);
944 	}
945 }
946 
947 /*
948  * The AIO daemon, most of the actual work is done in aio_process,
949  * but the setup (and address space mgmt) is done in this routine.
950  */
951 static void
952 aio_daemon(void *_id)
953 {
954 	struct aiocblist *aiocbe;
955 	struct aiothreadlist *aiop;
956 	struct kaioinfo *ki;
957 	struct proc *curcp, *mycp, *userp;
958 	struct vmspace *myvm, *tmpvm;
959 	struct thread *td = curthread;
960 	int id = (intptr_t)_id;
961 
962 	/*
963 	 * Local copies of curproc (cp) and vmspace (myvm)
964 	 */
965 	mycp = td->td_proc;
966 	myvm = mycp->p_vmspace;
967 
968 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
969 
970 	/*
971 	 * Allocate and ready the aio control info.  There is one aiop structure
972 	 * per daemon.
973 	 */
974 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
975 	aiop->aiothread = td;
976 	aiop->aiothreadflags = 0;
977 
978 	/* The daemon resides in its own pgrp. */
979 	setsid(td, NULL);
980 
981 	/*
982 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
983 	 * and creating too many daemons.)
984 	 */
985 	sema_post(&aio_newproc_sem);
986 
987 	mtx_lock(&aio_job_mtx);
988 	for (;;) {
989 		/*
990 		 * curcp is the current daemon process context.
991 		 * userp is the current user process context.
992 		 */
993 		curcp = mycp;
994 
995 		/*
996 		 * Take daemon off of free queue
997 		 */
998 		if (aiop->aiothreadflags & AIOP_FREE) {
999 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1000 			aiop->aiothreadflags &= ~AIOP_FREE;
1001 		}
1002 
1003 		/*
1004 		 * Check for jobs.
1005 		 */
1006 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1007 			mtx_unlock(&aio_job_mtx);
1008 			userp = aiocbe->userproc;
1009 
1010 			/*
1011 			 * Connect to process address space for user program.
1012 			 */
1013 			if (userp != curcp) {
1014 				/*
1015 				 * Save the current address space that we are
1016 				 * connected to.
1017 				 */
1018 				tmpvm = mycp->p_vmspace;
1019 
1020 				/*
1021 				 * Point to the new user address space, and
1022 				 * refer to it.
1023 				 */
1024 				mycp->p_vmspace = userp->p_vmspace;
1025 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1026 
1027 				/* Activate the new mapping. */
1028 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1029 
1030 				/*
1031 				 * If the old address space wasn't the daemons
1032 				 * own address space, then we need to remove the
1033 				 * daemon's reference from the other process
1034 				 * that it was acting on behalf of.
1035 				 */
1036 				if (tmpvm != myvm) {
1037 					vmspace_free(tmpvm);
1038 				}
1039 				curcp = userp;
1040 			}
1041 
1042 			ki = userp->p_aioinfo;
1043 
1044 			/* Do the I/O function. */
1045 			aio_process(aiocbe);
1046 
1047 			mtx_lock(&aio_job_mtx);
1048 			/* Decrement the active job count. */
1049 			ki->kaio_active_count--;
1050 			mtx_unlock(&aio_job_mtx);
1051 
1052 			AIO_LOCK(ki);
1053 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1054 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1055 			AIO_UNLOCK(ki);
1056 
1057 			mtx_lock(&aio_job_mtx);
1058 		}
1059 
1060 		/*
1061 		 * Disconnect from user address space.
1062 		 */
1063 		if (curcp != mycp) {
1064 
1065 			mtx_unlock(&aio_job_mtx);
1066 
1067 			/* Get the user address space to disconnect from. */
1068 			tmpvm = mycp->p_vmspace;
1069 
1070 			/* Get original address space for daemon. */
1071 			mycp->p_vmspace = myvm;
1072 
1073 			/* Activate the daemon's address space. */
1074 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1075 #ifdef DIAGNOSTIC
1076 			if (tmpvm == myvm) {
1077 				printf("AIOD: vmspace problem -- %d\n",
1078 				    mycp->p_pid);
1079 			}
1080 #endif
1081 			/* Remove our vmspace reference. */
1082 			vmspace_free(tmpvm);
1083 
1084 			curcp = mycp;
1085 
1086 			mtx_lock(&aio_job_mtx);
1087 			/*
1088 			 * We have to restart to avoid race, we only sleep if
1089 			 * no job can be selected, that should be
1090 			 * curcp == mycp.
1091 			 */
1092 			continue;
1093 		}
1094 
1095 		mtx_assert(&aio_job_mtx, MA_OWNED);
1096 
1097 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1098 		aiop->aiothreadflags |= AIOP_FREE;
1099 
1100 		/*
1101 		 * If daemon is inactive for a long time, allow it to exit,
1102 		 * thereby freeing resources.
1103 		 */
1104 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1105 		    aiod_lifetime)) {
1106 			if (TAILQ_EMPTY(&aio_jobs)) {
1107 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1108 				    (num_aio_procs > target_aio_procs)) {
1109 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1110 					num_aio_procs--;
1111 					mtx_unlock(&aio_job_mtx);
1112 					uma_zfree(aiop_zone, aiop);
1113 					free_unr(aiod_unr, id);
1114 #ifdef DIAGNOSTIC
1115 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1116 						printf("AIOD: bad vm refcnt for"
1117 						    " exiting daemon: %d\n",
1118 						    mycp->p_vmspace->vm_refcnt);
1119 					}
1120 #endif
1121 					kthread_exit(0);
1122 				}
1123 			}
1124 		}
1125 	}
1126 	mtx_unlock(&aio_job_mtx);
1127 	panic("shouldn't be here\n");
1128 }
1129 
1130 /*
1131  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1132  * AIO daemon modifies its environment itself.
1133  */
1134 static int
1135 aio_newproc(int *start)
1136 {
1137 	int error;
1138 	struct proc *p;
1139 	int id;
1140 
1141 	id = alloc_unr(aiod_unr);
1142 	error = kthread_create(aio_daemon, (void *)(intptr_t)id, &p,
1143 		RFNOWAIT, 0, "aiod%d", id);
1144 	if (error == 0) {
1145 		/*
1146 		 * Wait until daemon is started.
1147 		 */
1148 		sema_wait(&aio_newproc_sem);
1149 		mtx_lock(&aio_job_mtx);
1150 		num_aio_procs++;
1151 		if (start != NULL)
1152 			(*start)--;
1153 		mtx_unlock(&aio_job_mtx);
1154 	} else {
1155 		free_unr(aiod_unr, id);
1156 	}
1157 	return (error);
1158 }
1159 
1160 /*
1161  * Try the high-performance, low-overhead physio method for eligible
1162  * VCHR devices.  This method doesn't use an aio helper thread, and
1163  * thus has very low overhead.
1164  *
1165  * Assumes that the caller, aio_aqueue(), has incremented the file
1166  * structure's reference count, preventing its deallocation for the
1167  * duration of this call.
1168  */
1169 static int
1170 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1171 {
1172 	struct aiocb *cb;
1173 	struct file *fp;
1174 	struct buf *bp;
1175 	struct vnode *vp;
1176 	struct kaioinfo *ki;
1177 	struct aioliojob *lj;
1178 	int error;
1179 
1180 	cb = &aiocbe->uaiocb;
1181 	fp = aiocbe->fd_file;
1182 
1183 	if (fp->f_type != DTYPE_VNODE)
1184 		return (-1);
1185 
1186 	vp = fp->f_vnode;
1187 
1188 	/*
1189 	 * If its not a disk, we don't want to return a positive error.
1190 	 * It causes the aio code to not fall through to try the thread
1191 	 * way when you're talking to a regular file.
1192 	 */
1193 	if (!vn_isdisk(vp, &error)) {
1194 		if (error == ENOTBLK)
1195 			return (-1);
1196 		else
1197 			return (error);
1198 	}
1199 
1200 	if (vp->v_bufobj.bo_bsize == 0)
1201 		return (-1);
1202 
1203  	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1204 		return (-1);
1205 
1206 	if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1207 		return (-1);
1208 
1209 	if (cb->aio_nbytes >
1210 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1211 		return (-1);
1212 
1213 	ki = p->p_aioinfo;
1214 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1215 		return (-1);
1216 
1217 	/* Create and build a buffer header for a transfer. */
1218 	bp = (struct buf *)getpbuf(NULL);
1219 	BUF_KERNPROC(bp);
1220 
1221 	AIO_LOCK(ki);
1222 	ki->kaio_count++;
1223 	ki->kaio_buffer_count++;
1224 	lj = aiocbe->lio;
1225 	if (lj)
1226 		lj->lioj_count++;
1227 	AIO_UNLOCK(ki);
1228 
1229 	/*
1230 	 * Get a copy of the kva from the physical buffer.
1231 	 */
1232 	error = 0;
1233 
1234 	bp->b_bcount = cb->aio_nbytes;
1235 	bp->b_bufsize = cb->aio_nbytes;
1236 	bp->b_iodone = aio_physwakeup;
1237 	bp->b_saveaddr = bp->b_data;
1238 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1239 	bp->b_offset = cb->aio_offset;
1240 	bp->b_iooffset = cb->aio_offset;
1241 	bp->b_blkno = btodb(cb->aio_offset);
1242 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1243 
1244 	/*
1245 	 * Bring buffer into kernel space.
1246 	 */
1247 	if (vmapbuf(bp) < 0) {
1248 		error = EFAULT;
1249 		goto doerror;
1250 	}
1251 
1252 	AIO_LOCK(ki);
1253 	aiocbe->bp = bp;
1254 	bp->b_caller1 = (void *)aiocbe;
1255 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1256 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1257 	aiocbe->jobstate = JOBST_JOBQBUF;
1258 	cb->_aiocb_private.status = cb->aio_nbytes;
1259 	AIO_UNLOCK(ki);
1260 
1261 	atomic_add_int(&num_queue_count, 1);
1262 	atomic_add_int(&num_buf_aio, 1);
1263 
1264 	bp->b_error = 0;
1265 
1266 	TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1267 
1268 	/* Perform transfer. */
1269 	dev_strategy(vp->v_rdev, bp);
1270 	return (0);
1271 
1272 doerror:
1273 	AIO_LOCK(ki);
1274 	ki->kaio_count--;
1275 	ki->kaio_buffer_count--;
1276 	if (lj)
1277 		lj->lioj_count--;
1278 	aiocbe->bp = NULL;
1279 	AIO_UNLOCK(ki);
1280 	relpbuf(bp, NULL);
1281 	return (error);
1282 }
1283 
1284 /*
1285  * Wake up aio requests that may be serviceable now.
1286  */
1287 static void
1288 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1289 {
1290 	struct aiocblist *cb, *cbn;
1291 	int opcode;
1292 
1293 	if (sb == &so->so_snd)
1294 		opcode = LIO_WRITE;
1295 	else
1296 		opcode = LIO_READ;
1297 
1298 	SOCKBUF_LOCK(sb);
1299 	sb->sb_flags &= ~SB_AIO;
1300 	mtx_lock(&aio_job_mtx);
1301 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1302 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1303 			if (cb->jobstate != JOBST_JOBQSOCK)
1304 				panic("invalid queue value");
1305 			/* XXX
1306 			 * We don't have actual sockets backend yet,
1307 			 * so we simply move the requests to the generic
1308 			 * file I/O backend.
1309 			 */
1310 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1311 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1312 			aio_kick_nowait(cb->userproc);
1313 		}
1314 	}
1315 	mtx_unlock(&aio_job_mtx);
1316 	SOCKBUF_UNLOCK(sb);
1317 }
1318 
1319 /*
1320  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1321  * technique is done in this code.
1322  */
1323 int
1324 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1325 	int type, int oldsigev)
1326 {
1327 	struct proc *p = td->td_proc;
1328 	struct file *fp;
1329 	struct socket *so;
1330 	struct aiocblist *aiocbe, *cb;
1331 	struct kaioinfo *ki;
1332 	struct kevent kev;
1333 	struct sockbuf *sb;
1334 	int opcode;
1335 	int error;
1336 	int fd, kqfd;
1337 	int jid;
1338 
1339 	if (p->p_aioinfo == NULL)
1340 		aio_init_aioinfo(p);
1341 
1342 	ki = p->p_aioinfo;
1343 
1344 	suword(&job->_aiocb_private.status, -1);
1345 	suword(&job->_aiocb_private.error, 0);
1346 	suword(&job->_aiocb_private.kernelinfo, -1);
1347 
1348 	if (num_queue_count >= max_queue_count ||
1349 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1350 		suword(&job->_aiocb_private.error, EAGAIN);
1351 		return (EAGAIN);
1352 	}
1353 
1354 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1355 	aiocbe->inputcharge = 0;
1356 	aiocbe->outputcharge = 0;
1357 	knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL);
1358 
1359 	if (oldsigev) {
1360 		bzero(&aiocbe->uaiocb, sizeof(struct aiocb));
1361 		error = copyin(job, &aiocbe->uaiocb, sizeof(struct oaiocb));
1362 		bcopy(&aiocbe->uaiocb.__spare__, &aiocbe->uaiocb.aio_sigevent,
1363 			sizeof(struct osigevent));
1364 	} else {
1365 		error = copyin(job, &aiocbe->uaiocb, sizeof(struct aiocb));
1366 	}
1367 	if (error) {
1368 		suword(&job->_aiocb_private.error, error);
1369 		uma_zfree(aiocb_zone, aiocbe);
1370 		return (error);
1371 	}
1372 
1373 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1374 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1375 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1376 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1377 		suword(&job->_aiocb_private.error, EINVAL);
1378 		uma_zfree(aiocb_zone, aiocbe);
1379 		return (EINVAL);
1380 	}
1381 
1382 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1383 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1384 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1385 		uma_zfree(aiocb_zone, aiocbe);
1386 		return (EINVAL);
1387 	}
1388 
1389 	ksiginfo_init(&aiocbe->ksi);
1390 
1391 	/* Save userspace address of the job info. */
1392 	aiocbe->uuaiocb = job;
1393 
1394 	/* Get the opcode. */
1395 	if (type != LIO_NOP)
1396 		aiocbe->uaiocb.aio_lio_opcode = type;
1397 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1398 
1399 	/* Fetch the file object for the specified file descriptor. */
1400 	fd = aiocbe->uaiocb.aio_fildes;
1401 	switch (opcode) {
1402 	case LIO_WRITE:
1403 		error = fget_write(td, fd, &fp);
1404 		break;
1405 	case LIO_READ:
1406 		error = fget_read(td, fd, &fp);
1407 		break;
1408 	default:
1409 		error = fget(td, fd, &fp);
1410 	}
1411 	if (error) {
1412 		uma_zfree(aiocb_zone, aiocbe);
1413 		suword(&job->_aiocb_private.error, error);
1414 		return (error);
1415 	}
1416 
1417 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1418 		error = EINVAL;
1419 		goto aqueue_fail;
1420 	}
1421 
1422 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1423 		error = EINVAL;
1424 		goto aqueue_fail;
1425 	}
1426 
1427 	aiocbe->fd_file = fp;
1428 
1429 	mtx_lock(&aio_job_mtx);
1430 	jid = jobrefid++;
1431 	aiocbe->seqno = jobseqno++;
1432 	mtx_unlock(&aio_job_mtx);
1433 	error = suword(&job->_aiocb_private.kernelinfo, jid);
1434 	if (error) {
1435 		error = EINVAL;
1436 		goto aqueue_fail;
1437 	}
1438 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1439 
1440 	if (opcode == LIO_NOP) {
1441 		fdrop(fp, td);
1442 		uma_zfree(aiocb_zone, aiocbe);
1443 		return (0);
1444 	}
1445 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1446 	    (opcode != LIO_SYNC)) {
1447 		error = EINVAL;
1448 		goto aqueue_fail;
1449 	}
1450 
1451 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1452 		goto no_kqueue;
1453 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1454 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1455 	kev.filter = EVFILT_AIO;
1456 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1457 	kev.data = (intptr_t)aiocbe;
1458 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1459 	error = kqfd_register(kqfd, &kev, td, 1);
1460 aqueue_fail:
1461 	if (error) {
1462 		fdrop(fp, td);
1463 		uma_zfree(aiocb_zone, aiocbe);
1464 		suword(&job->_aiocb_private.error, error);
1465 		goto done;
1466 	}
1467 no_kqueue:
1468 
1469 	suword(&job->_aiocb_private.error, EINPROGRESS);
1470 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1471 	aiocbe->userproc = p;
1472 	aiocbe->cred = crhold(td->td_ucred);
1473 	aiocbe->jobflags = 0;
1474 	aiocbe->lio = lj;
1475 
1476 	if (opcode == LIO_SYNC)
1477 		goto queueit;
1478 
1479 	if (fp->f_type == DTYPE_SOCKET) {
1480 		/*
1481 		 * Alternate queueing for socket ops: Reach down into the
1482 		 * descriptor to get the socket data.  Then check to see if the
1483 		 * socket is ready to be read or written (based on the requested
1484 		 * operation).
1485 		 *
1486 		 * If it is not ready for io, then queue the aiocbe on the
1487 		 * socket, and set the flags so we get a call when sbnotify()
1488 		 * happens.
1489 		 *
1490 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1491 		 * and unlock the snd sockbuf for no reason.
1492 		 */
1493 		so = fp->f_data;
1494 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1495 		SOCKBUF_LOCK(sb);
1496 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1497 		    LIO_WRITE) && (!sowriteable(so)))) {
1498 			sb->sb_flags |= SB_AIO;
1499 
1500 			mtx_lock(&aio_job_mtx);
1501 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1502 			mtx_unlock(&aio_job_mtx);
1503 
1504 			AIO_LOCK(ki);
1505 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1506 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1507 			aiocbe->jobstate = JOBST_JOBQSOCK;
1508 			ki->kaio_count++;
1509 			if (lj)
1510 				lj->lioj_count++;
1511 			AIO_UNLOCK(ki);
1512 			SOCKBUF_UNLOCK(sb);
1513 			atomic_add_int(&num_queue_count, 1);
1514 			error = 0;
1515 			goto done;
1516 		}
1517 		SOCKBUF_UNLOCK(sb);
1518 	}
1519 
1520 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1521 		goto done;
1522 #if 0
1523 	if (error > 0) {
1524 		aiocbe->uaiocb._aiocb_private.error = error;
1525 		suword(&job->_aiocb_private.error, error);
1526 		goto done;
1527 	}
1528 #endif
1529 queueit:
1530 	/* No buffer for daemon I/O. */
1531 	aiocbe->bp = NULL;
1532 	atomic_add_int(&num_queue_count, 1);
1533 
1534 	AIO_LOCK(ki);
1535 	ki->kaio_count++;
1536 	if (lj)
1537 		lj->lioj_count++;
1538 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1539 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1540 	if (opcode == LIO_SYNC) {
1541 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1542 			if (cb->fd_file == aiocbe->fd_file &&
1543 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1544 			    cb->seqno < aiocbe->seqno) {
1545 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1546 				aiocbe->pending++;
1547 			}
1548 		}
1549 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1550 			if (cb->fd_file == aiocbe->fd_file &&
1551 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1552 			    cb->seqno < aiocbe->seqno) {
1553 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1554 				aiocbe->pending++;
1555 			}
1556 		}
1557 		if (aiocbe->pending != 0) {
1558 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1559 			aiocbe->jobstate = JOBST_JOBQSYNC;
1560 			AIO_UNLOCK(ki);
1561 			goto done;
1562 		}
1563 	}
1564 	mtx_lock(&aio_job_mtx);
1565 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1566 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1567 	aio_kick_nowait(p);
1568 	mtx_unlock(&aio_job_mtx);
1569 	AIO_UNLOCK(ki);
1570 	error = 0;
1571 done:
1572 	return (error);
1573 }
1574 
1575 static void
1576 aio_kick_nowait(struct proc *userp)
1577 {
1578 	struct kaioinfo *ki = userp->p_aioinfo;
1579 	struct aiothreadlist *aiop;
1580 
1581 	mtx_assert(&aio_job_mtx, MA_OWNED);
1582 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1583 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1584 		aiop->aiothreadflags &= ~AIOP_FREE;
1585 		wakeup(aiop->aiothread);
1586 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1587 	    ((ki->kaio_active_count + num_aio_resv_start) <
1588 	    ki->kaio_maxactive_count)) {
1589 		taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1590 	}
1591 }
1592 
1593 static int
1594 aio_kick(struct proc *userp)
1595 {
1596 	struct kaioinfo *ki = userp->p_aioinfo;
1597 	struct aiothreadlist *aiop;
1598 	int error, ret = 0;
1599 
1600 	mtx_assert(&aio_job_mtx, MA_OWNED);
1601 retryproc:
1602 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1603 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1604 		aiop->aiothreadflags &= ~AIOP_FREE;
1605 		wakeup(aiop->aiothread);
1606 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1607 	    ((ki->kaio_active_count + num_aio_resv_start) <
1608 	    ki->kaio_maxactive_count)) {
1609 		num_aio_resv_start++;
1610 		mtx_unlock(&aio_job_mtx);
1611 		error = aio_newproc(&num_aio_resv_start);
1612 		mtx_lock(&aio_job_mtx);
1613 		if (error) {
1614 			num_aio_resv_start--;
1615 			goto retryproc;
1616 		}
1617 	} else {
1618 		ret = -1;
1619 	}
1620 	return (ret);
1621 }
1622 
1623 static void
1624 aio_kick_helper(void *context, int pending)
1625 {
1626 	struct proc *userp = context;
1627 
1628 	mtx_lock(&aio_job_mtx);
1629 	while (--pending >= 0) {
1630 		if (aio_kick(userp))
1631 			break;
1632 	}
1633 	mtx_unlock(&aio_job_mtx);
1634 }
1635 
1636 /*
1637  * Support the aio_return system call, as a side-effect, kernel resources are
1638  * released.
1639  */
1640 int
1641 aio_return(struct thread *td, struct aio_return_args *uap)
1642 {
1643 	struct proc *p = td->td_proc;
1644 	struct aiocblist *cb;
1645 	struct aiocb *uaiocb;
1646 	struct kaioinfo *ki;
1647 	int status, error;
1648 
1649 	ki = p->p_aioinfo;
1650 	if (ki == NULL)
1651 		return (EINVAL);
1652 	uaiocb = uap->aiocbp;
1653 	AIO_LOCK(ki);
1654 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1655 		if (cb->uuaiocb == uaiocb)
1656 			break;
1657 	}
1658 	if (cb != NULL) {
1659 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1660 		status = cb->uaiocb._aiocb_private.status;
1661 		error = cb->uaiocb._aiocb_private.error;
1662 		td->td_retval[0] = status;
1663 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1664 			td->td_ru.ru_oublock += cb->outputcharge;
1665 			cb->outputcharge = 0;
1666 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1667 			td->td_ru.ru_inblock += cb->inputcharge;
1668 			cb->inputcharge = 0;
1669 		}
1670 		aio_free_entry(cb);
1671 		AIO_UNLOCK(ki);
1672 		suword(&uaiocb->_aiocb_private.error, error);
1673 		suword(&uaiocb->_aiocb_private.status, status);
1674 	} else {
1675 		error = EINVAL;
1676 		AIO_UNLOCK(ki);
1677 	}
1678 	return (error);
1679 }
1680 
1681 /*
1682  * Allow a process to wakeup when any of the I/O requests are completed.
1683  */
1684 int
1685 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1686 {
1687 	struct proc *p = td->td_proc;
1688 	struct timeval atv;
1689 	struct timespec ts;
1690 	struct aiocb *const *cbptr, *cbp;
1691 	struct kaioinfo *ki;
1692 	struct aiocblist *cb, *cbfirst;
1693 	struct aiocb **ujoblist;
1694 	int njoblist;
1695 	int error;
1696 	int timo;
1697 	int i;
1698 
1699 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1700 		return (EINVAL);
1701 
1702 	timo = 0;
1703 	if (uap->timeout) {
1704 		/* Get timespec struct. */
1705 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1706 			return (error);
1707 
1708 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1709 			return (EINVAL);
1710 
1711 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1712 		if (itimerfix(&atv))
1713 			return (EINVAL);
1714 		timo = tvtohz(&atv);
1715 	}
1716 
1717 	ki = p->p_aioinfo;
1718 	if (ki == NULL)
1719 		return (EAGAIN);
1720 
1721 	njoblist = 0;
1722 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1723 	cbptr = uap->aiocbp;
1724 
1725 	for (i = 0; i < uap->nent; i++) {
1726 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1727 		if (cbp == 0)
1728 			continue;
1729 		ujoblist[njoblist] = cbp;
1730 		njoblist++;
1731 	}
1732 
1733 	if (njoblist == 0) {
1734 		uma_zfree(aiol_zone, ujoblist);
1735 		return (0);
1736 	}
1737 
1738 	AIO_LOCK(ki);
1739 	for (;;) {
1740 		cbfirst = NULL;
1741 		error = 0;
1742 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1743 			for (i = 0; i < njoblist; i++) {
1744 				if (cb->uuaiocb == ujoblist[i]) {
1745 					if (cbfirst == NULL)
1746 						cbfirst = cb;
1747 					if (cb->jobstate == JOBST_JOBFINISHED)
1748 						goto RETURN;
1749 				}
1750 			}
1751 		}
1752 		/* All tasks were finished. */
1753 		if (cbfirst == NULL)
1754 			break;
1755 
1756 		ki->kaio_flags |= KAIO_WAKEUP;
1757 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1758 		    "aiospn", timo);
1759 		if (error == ERESTART)
1760 			error = EINTR;
1761 		if (error)
1762 			break;
1763 	}
1764 RETURN:
1765 	AIO_UNLOCK(ki);
1766 	uma_zfree(aiol_zone, ujoblist);
1767 	return (error);
1768 }
1769 
1770 /*
1771  * aio_cancel cancels any non-physio aio operations not currently in
1772  * progress.
1773  */
1774 int
1775 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1776 {
1777 	struct proc *p = td->td_proc;
1778 	struct kaioinfo *ki;
1779 	struct aiocblist *cbe, *cbn;
1780 	struct file *fp;
1781 	struct socket *so;
1782 	int error;
1783 	int remove;
1784 	int cancelled = 0;
1785 	int notcancelled = 0;
1786 	struct vnode *vp;
1787 
1788 	/* Lookup file object. */
1789 	error = fget(td, uap->fd, &fp);
1790 	if (error)
1791 		return (error);
1792 
1793 	ki = p->p_aioinfo;
1794 	if (ki == NULL)
1795 		goto done;
1796 
1797 	if (fp->f_type == DTYPE_VNODE) {
1798 		vp = fp->f_vnode;
1799 		if (vn_isdisk(vp, &error)) {
1800 			fdrop(fp, td);
1801 			td->td_retval[0] = AIO_NOTCANCELED;
1802 			return (0);
1803 		}
1804 	}
1805 
1806 	AIO_LOCK(ki);
1807 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1808 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1809 		    ((uap->aiocbp == NULL) ||
1810 		     (uap->aiocbp == cbe->uuaiocb))) {
1811 			remove = 0;
1812 
1813 			mtx_lock(&aio_job_mtx);
1814 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1815 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1816 				remove = 1;
1817 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
1818 				MPASS(fp->f_type == DTYPE_SOCKET);
1819 				so = fp->f_data;
1820 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1821 				remove = 1;
1822 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
1823 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1824 				remove = 1;
1825 			}
1826 			mtx_unlock(&aio_job_mtx);
1827 
1828 			if (remove) {
1829 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1830 				cbe->uaiocb._aiocb_private.status = -1;
1831 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1832 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
1833 				cancelled++;
1834 			} else {
1835 				notcancelled++;
1836 			}
1837 			if (uap->aiocbp != NULL)
1838 				break;
1839 		}
1840 	}
1841 	AIO_UNLOCK(ki);
1842 
1843 done:
1844 	fdrop(fp, td);
1845 
1846 	if (uap->aiocbp != NULL) {
1847 		if (cancelled) {
1848 			td->td_retval[0] = AIO_CANCELED;
1849 			return (0);
1850 		}
1851 	}
1852 
1853 	if (notcancelled) {
1854 		td->td_retval[0] = AIO_NOTCANCELED;
1855 		return (0);
1856 	}
1857 
1858 	if (cancelled) {
1859 		td->td_retval[0] = AIO_CANCELED;
1860 		return (0);
1861 	}
1862 
1863 	td->td_retval[0] = AIO_ALLDONE;
1864 
1865 	return (0);
1866 }
1867 
1868 /*
1869  * aio_error is implemented in the kernel level for compatibility purposes
1870  * only.  For a user mode async implementation, it would be best to do it in
1871  * a userland subroutine.
1872  */
1873 int
1874 aio_error(struct thread *td, struct aio_error_args *uap)
1875 {
1876 	struct proc *p = td->td_proc;
1877 	struct aiocblist *cb;
1878 	struct kaioinfo *ki;
1879 	int status;
1880 
1881 	ki = p->p_aioinfo;
1882 	if (ki == NULL) {
1883 		td->td_retval[0] = EINVAL;
1884 		return (0);
1885 	}
1886 
1887 	AIO_LOCK(ki);
1888 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1889 		if (cb->uuaiocb == uap->aiocbp) {
1890 			if (cb->jobstate == JOBST_JOBFINISHED)
1891 				td->td_retval[0] =
1892 					cb->uaiocb._aiocb_private.error;
1893 			else
1894 				td->td_retval[0] = EINPROGRESS;
1895 			AIO_UNLOCK(ki);
1896 			return (0);
1897 		}
1898 	}
1899 	AIO_UNLOCK(ki);
1900 
1901 	/*
1902 	 * Hack for failure of aio_aqueue.
1903 	 */
1904 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1905 	if (status == -1) {
1906 		td->td_retval[0] = fuword(&uap->aiocbp->_aiocb_private.error);
1907 		return (0);
1908 	}
1909 
1910 	td->td_retval[0] = EINVAL;
1911 	return (0);
1912 }
1913 
1914 /* syscall - asynchronous read from a file (REALTIME) */
1915 int
1916 oaio_read(struct thread *td, struct oaio_read_args *uap)
1917 {
1918 
1919 	return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 1);
1920 }
1921 
1922 int
1923 aio_read(struct thread *td, struct aio_read_args *uap)
1924 {
1925 
1926 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, 0);
1927 }
1928 
1929 /* syscall - asynchronous write to a file (REALTIME) */
1930 int
1931 oaio_write(struct thread *td, struct oaio_write_args *uap)
1932 {
1933 
1934 	return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 1);
1935 }
1936 
1937 int
1938 aio_write(struct thread *td, struct aio_write_args *uap)
1939 {
1940 
1941 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, 0);
1942 }
1943 
1944 /* syscall - list directed I/O (REALTIME) */
1945 int
1946 olio_listio(struct thread *td, struct olio_listio_args *uap)
1947 {
1948 	return do_lio_listio(td, (struct lio_listio_args *)uap, 1);
1949 }
1950 
1951 /* syscall - list directed I/O (REALTIME) */
1952 int
1953 lio_listio(struct thread *td, struct lio_listio_args *uap)
1954 {
1955 	return do_lio_listio(td, uap, 0);
1956 }
1957 
1958 static int
1959 do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev)
1960 {
1961 	struct proc *p = td->td_proc;
1962 	struct aiocb *iocb, * const *cbptr;
1963 	struct kaioinfo *ki;
1964 	struct aioliojob *lj;
1965 	struct kevent kev;
1966 	int nent;
1967 	int error;
1968 	int nerror;
1969 	int i;
1970 
1971 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1972 		return (EINVAL);
1973 
1974 	nent = uap->nent;
1975 	if (nent < 0 || nent > AIO_LISTIO_MAX)
1976 		return (EINVAL);
1977 
1978 	if (p->p_aioinfo == NULL)
1979 		aio_init_aioinfo(p);
1980 
1981 	ki = p->p_aioinfo;
1982 
1983 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
1984 	lj->lioj_flags = 0;
1985 	lj->lioj_count = 0;
1986 	lj->lioj_finished_count = 0;
1987 	knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL);
1988 	ksiginfo_init(&lj->lioj_ksi);
1989 
1990 	/*
1991 	 * Setup signal.
1992 	 */
1993 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1994 		bzero(&lj->lioj_signal, sizeof(&lj->lioj_signal));
1995 		error = copyin(uap->sig, &lj->lioj_signal,
1996 				oldsigev ? sizeof(struct osigevent) :
1997 					   sizeof(struct sigevent));
1998 		if (error) {
1999 			uma_zfree(aiolio_zone, lj);
2000 			return (error);
2001 		}
2002 
2003 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2004 			/* Assume only new style KEVENT */
2005 			kev.filter = EVFILT_LIO;
2006 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2007 			kev.ident = (uintptr_t)uap->acb_list; /* something unique */
2008 			kev.data = (intptr_t)lj;
2009 			/* pass user defined sigval data */
2010 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2011 			error = kqfd_register(
2012 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2013 			if (error) {
2014 				uma_zfree(aiolio_zone, lj);
2015 				return (error);
2016 			}
2017 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2018 			;
2019 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2020 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2021 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2022 					uma_zfree(aiolio_zone, lj);
2023 					return EINVAL;
2024 				}
2025 				lj->lioj_flags |= LIOJ_SIGNAL;
2026 		} else {
2027 			uma_zfree(aiolio_zone, lj);
2028 			return EINVAL;
2029 		}
2030 	}
2031 
2032 	AIO_LOCK(ki);
2033 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2034 	/*
2035 	 * Add extra aiocb count to avoid the lio to be freed
2036 	 * by other threads doing aio_waitcomplete or aio_return,
2037 	 * and prevent event from being sent until we have queued
2038 	 * all tasks.
2039 	 */
2040 	lj->lioj_count = 1;
2041 	AIO_UNLOCK(ki);
2042 
2043 	/*
2044 	 * Get pointers to the list of I/O requests.
2045 	 */
2046 	nerror = 0;
2047 	cbptr = uap->acb_list;
2048 	for (i = 0; i < uap->nent; i++) {
2049 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2050 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2051 			error = aio_aqueue(td, iocb, lj, LIO_NOP, oldsigev);
2052 			if (error != 0)
2053 				nerror++;
2054 		}
2055 	}
2056 
2057 	error = 0;
2058 	AIO_LOCK(ki);
2059 	if (uap->mode == LIO_WAIT) {
2060 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2061 			ki->kaio_flags |= KAIO_WAKEUP;
2062 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2063 			    PRIBIO | PCATCH, "aiospn", 0);
2064 			if (error == ERESTART)
2065 				error = EINTR;
2066 			if (error)
2067 				break;
2068 		}
2069 	} else {
2070 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2071 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2072 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2073 				KNOTE_LOCKED(&lj->klist, 1);
2074 			}
2075 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2076 			    == LIOJ_SIGNAL
2077 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2078 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2079 				aio_sendsig(p, &lj->lioj_signal,
2080 					    &lj->lioj_ksi);
2081 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2082 			}
2083 		}
2084 	}
2085 	lj->lioj_count--;
2086 	if (lj->lioj_count == 0) {
2087 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2088 		knlist_delete(&lj->klist, curthread, 1);
2089 		PROC_LOCK(p);
2090 		sigqueue_take(&lj->lioj_ksi);
2091 		PROC_UNLOCK(p);
2092 		AIO_UNLOCK(ki);
2093 		uma_zfree(aiolio_zone, lj);
2094 	} else
2095 		AIO_UNLOCK(ki);
2096 
2097 	if (nerror)
2098 		return (EIO);
2099 	return (error);
2100 }
2101 
2102 /*
2103  * Called from interrupt thread for physio, we should return as fast
2104  * as possible, so we schedule a biohelper task.
2105  */
2106 static void
2107 aio_physwakeup(struct buf *bp)
2108 {
2109 	struct aiocblist *aiocbe;
2110 
2111 	aiocbe = (struct aiocblist *)bp->b_caller1;
2112 	taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2113 }
2114 
2115 /*
2116  * Task routine to perform heavy tasks, process wakeup, and signals.
2117  */
2118 static void
2119 biohelper(void *context, int pending)
2120 {
2121 	struct aiocblist *aiocbe = context;
2122 	struct buf *bp;
2123 	struct proc *userp;
2124 	struct kaioinfo *ki;
2125 	int nblks;
2126 
2127 	bp = aiocbe->bp;
2128 	userp = aiocbe->userproc;
2129 	ki = userp->p_aioinfo;
2130 	AIO_LOCK(ki);
2131 	aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2132 	aiocbe->uaiocb._aiocb_private.error = 0;
2133 	if (bp->b_ioflags & BIO_ERROR)
2134 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2135 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2136 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2137 		aiocbe->outputcharge += nblks;
2138 	else
2139 		aiocbe->inputcharge += nblks;
2140 	aiocbe->bp = NULL;
2141 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2142 	ki->kaio_buffer_count--;
2143 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2144 	AIO_UNLOCK(ki);
2145 
2146 	/* Release mapping into kernel space. */
2147 	vunmapbuf(bp);
2148 	relpbuf(bp, NULL);
2149 	atomic_subtract_int(&num_buf_aio, 1);
2150 }
2151 
2152 /* syscall - wait for the next completion of an aio request */
2153 int
2154 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2155 {
2156 	struct proc *p = td->td_proc;
2157 	struct timeval atv;
2158 	struct timespec ts;
2159 	struct kaioinfo *ki;
2160 	struct aiocblist *cb;
2161 	struct aiocb *uuaiocb;
2162 	int error, status, timo;
2163 
2164 	suword(uap->aiocbp, (long)NULL);
2165 
2166 	timo = 0;
2167 	if (uap->timeout) {
2168 		/* Get timespec struct. */
2169 		error = copyin(uap->timeout, &ts, sizeof(ts));
2170 		if (error)
2171 			return (error);
2172 
2173 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2174 			return (EINVAL);
2175 
2176 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2177 		if (itimerfix(&atv))
2178 			return (EINVAL);
2179 		timo = tvtohz(&atv);
2180 	}
2181 
2182 	if (p->p_aioinfo == NULL)
2183 		aio_init_aioinfo(p);
2184 	ki = p->p_aioinfo;
2185 
2186 	error = 0;
2187 	cb = NULL;
2188 	AIO_LOCK(ki);
2189 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2190 		ki->kaio_flags |= KAIO_WAKEUP;
2191 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2192 		    "aiowc", timo);
2193 		if (timo && error == ERESTART)
2194 			error = EINTR;
2195 		if (error)
2196 			break;
2197 	}
2198 
2199 	if (cb != NULL) {
2200 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2201 		uuaiocb = cb->uuaiocb;
2202 		status = cb->uaiocb._aiocb_private.status;
2203 		error = cb->uaiocb._aiocb_private.error;
2204 		td->td_retval[0] = status;
2205 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2206 			td->td_ru.ru_oublock += cb->outputcharge;
2207 			cb->outputcharge = 0;
2208 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2209 			td->td_ru.ru_inblock += cb->inputcharge;
2210 			cb->inputcharge = 0;
2211 		}
2212 		aio_free_entry(cb);
2213 		AIO_UNLOCK(ki);
2214 		suword(uap->aiocbp, (long)uuaiocb);
2215 		suword(&uuaiocb->_aiocb_private.error, error);
2216 		suword(&uuaiocb->_aiocb_private.status, status);
2217 	} else
2218 		AIO_UNLOCK(ki);
2219 
2220 	return (error);
2221 }
2222 
2223 int
2224 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2225 {
2226 	struct proc *p = td->td_proc;
2227 	struct kaioinfo *ki;
2228 
2229 	if (uap->op != O_SYNC) /* XXX lack of O_DSYNC */
2230 		return (EINVAL);
2231 	ki = p->p_aioinfo;
2232 	if (ki == NULL)
2233 		aio_init_aioinfo(p);
2234 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_SYNC, 0);
2235 }
2236 
2237 /* kqueue attach function */
2238 static int
2239 filt_aioattach(struct knote *kn)
2240 {
2241 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2242 
2243 	/*
2244 	 * The aiocbe pointer must be validated before using it, so
2245 	 * registration is restricted to the kernel; the user cannot
2246 	 * set EV_FLAG1.
2247 	 */
2248 	if ((kn->kn_flags & EV_FLAG1) == 0)
2249 		return (EPERM);
2250 	kn->kn_flags &= ~EV_FLAG1;
2251 
2252 	knlist_add(&aiocbe->klist, kn, 0);
2253 
2254 	return (0);
2255 }
2256 
2257 /* kqueue detach function */
2258 static void
2259 filt_aiodetach(struct knote *kn)
2260 {
2261 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2262 
2263 	if (!knlist_empty(&aiocbe->klist))
2264 		knlist_remove(&aiocbe->klist, kn, 0);
2265 }
2266 
2267 /* kqueue filter function */
2268 /*ARGSUSED*/
2269 static int
2270 filt_aio(struct knote *kn, long hint)
2271 {
2272 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2273 
2274 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2275 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2276 		return (0);
2277 	kn->kn_flags |= EV_EOF;
2278 	return (1);
2279 }
2280 
2281 /* kqueue attach function */
2282 static int
2283 filt_lioattach(struct knote *kn)
2284 {
2285 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2286 
2287 	/*
2288 	 * The aioliojob pointer must be validated before using it, so
2289 	 * registration is restricted to the kernel; the user cannot
2290 	 * set EV_FLAG1.
2291 	 */
2292 	if ((kn->kn_flags & EV_FLAG1) == 0)
2293 		return (EPERM);
2294 	kn->kn_flags &= ~EV_FLAG1;
2295 
2296 	knlist_add(&lj->klist, kn, 0);
2297 
2298 	return (0);
2299 }
2300 
2301 /* kqueue detach function */
2302 static void
2303 filt_liodetach(struct knote *kn)
2304 {
2305 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2306 
2307 	if (!knlist_empty(&lj->klist))
2308 		knlist_remove(&lj->klist, kn, 0);
2309 }
2310 
2311 /* kqueue filter function */
2312 /*ARGSUSED*/
2313 static int
2314 filt_lio(struct knote *kn, long hint)
2315 {
2316 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2317 
2318 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2319 }
2320