xref: /freebsd/sys/kern/vfs_aio.c (revision f3bb407b7c1b3faa88d0580541f01a8e6fb6cc68)
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 proc *mycp;
795 	struct aiocb *cb;
796 	struct file *fp;
797 	struct socket *so;
798 	struct uio auio;
799 	struct iovec aiov;
800 	int cnt;
801 	int error;
802 	int oublock_st, oublock_end;
803 	int inblock_st, inblock_end;
804 
805 	td = curthread;
806 	td_savedcred = td->td_ucred;
807 	td->td_ucred = aiocbe->cred;
808 	mycp = td->td_proc;
809 	cb = &aiocbe->uaiocb;
810 	fp = aiocbe->fd_file;
811 
812 	if (cb->aio_lio_opcode == LIO_SYNC) {
813 		error = 0;
814 		cnt = 0;
815 		if (fp->f_vnode != NULL)
816 			error = aio_fsync_vnode(td, fp->f_vnode);
817 		cb->_aiocb_private.error = error;
818 		cb->_aiocb_private.status = 0;
819 		td->td_ucred = td_savedcred;
820 		return;
821 	}
822 
823 	aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
824 	aiov.iov_len = cb->aio_nbytes;
825 
826 	auio.uio_iov = &aiov;
827 	auio.uio_iovcnt = 1;
828 	auio.uio_offset = cb->aio_offset;
829 	auio.uio_resid = cb->aio_nbytes;
830 	cnt = cb->aio_nbytes;
831 	auio.uio_segflg = UIO_USERSPACE;
832 	auio.uio_td = td;
833 
834 	inblock_st = mycp->p_stats->p_ru.ru_inblock;
835 	oublock_st = mycp->p_stats->p_ru.ru_oublock;
836 	/*
837 	 * aio_aqueue() acquires a reference to the file that is
838 	 * released in aio_free_entry().
839 	 */
840 	if (cb->aio_lio_opcode == LIO_READ) {
841 		auio.uio_rw = UIO_READ;
842 		error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, td);
843 	} else {
844 		if (fp->f_type == DTYPE_VNODE)
845 			bwillwrite();
846 		auio.uio_rw = UIO_WRITE;
847 		error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, td);
848 	}
849 	inblock_end = mycp->p_stats->p_ru.ru_inblock;
850 	oublock_end = mycp->p_stats->p_ru.ru_oublock;
851 
852 	aiocbe->inputcharge = inblock_end - inblock_st;
853 	aiocbe->outputcharge = oublock_end - oublock_st;
854 
855 	if ((error) && (auio.uio_resid != cnt)) {
856 		if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
857 			error = 0;
858 		if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE)) {
859 			int sigpipe = 1;
860 			if (fp->f_type == DTYPE_SOCKET) {
861 				so = fp->f_data;
862 				if (so->so_options & SO_NOSIGPIPE)
863 					sigpipe = 0;
864 			}
865 			if (sigpipe) {
866 				PROC_LOCK(aiocbe->userproc);
867 				psignal(aiocbe->userproc, SIGPIPE);
868 				PROC_UNLOCK(aiocbe->userproc);
869 			}
870 		}
871 	}
872 
873 	cnt -= auio.uio_resid;
874 	cb->_aiocb_private.error = error;
875 	cb->_aiocb_private.status = cnt;
876 	td->td_ucred = td_savedcred;
877 }
878 
879 static void
880 aio_bio_done_notify(struct proc *userp, struct aiocblist *aiocbe, int type)
881 {
882 	struct aioliojob *lj;
883 	struct kaioinfo *ki;
884 	struct aiocblist *scb, *scbn;
885 	int lj_done;
886 
887 	ki = userp->p_aioinfo;
888 	AIO_LOCK_ASSERT(ki, MA_OWNED);
889 	lj = aiocbe->lio;
890 	lj_done = 0;
891 	if (lj) {
892 		lj->lioj_finished_count++;
893 		if (lj->lioj_count == lj->lioj_finished_count)
894 			lj_done = 1;
895 	}
896 	if (type == DONE_QUEUE) {
897 		aiocbe->jobflags |= AIOCBLIST_DONE;
898 	} else {
899 		aiocbe->jobflags |= AIOCBLIST_BUFDONE;
900 	}
901 	TAILQ_INSERT_TAIL(&ki->kaio_done, aiocbe, plist);
902 	aiocbe->jobstate = JOBST_JOBFINISHED;
903 
904 	if (ki->kaio_flags & KAIO_RUNDOWN)
905 		goto notification_done;
906 
907 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
908 	    aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID)
909 		aio_sendsig(userp, &aiocbe->uaiocb.aio_sigevent, &aiocbe->ksi);
910 
911 	KNOTE_LOCKED(&aiocbe->klist, 1);
912 
913 	if (lj_done) {
914 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
915 			lj->lioj_flags |= LIOJ_KEVENT_POSTED;
916 			KNOTE_LOCKED(&lj->klist, 1);
917 		}
918 		if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
919 		    == LIOJ_SIGNAL
920 		    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
921 		        lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
922 			aio_sendsig(userp, &lj->lioj_signal, &lj->lioj_ksi);
923 			lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
924 		}
925 	}
926 
927 notification_done:
928 	if (aiocbe->jobflags & AIOCBLIST_CHECKSYNC) {
929 		TAILQ_FOREACH_SAFE(scb, &ki->kaio_syncqueue, list, scbn) {
930 			if (aiocbe->fd_file == scb->fd_file &&
931 			    aiocbe->seqno < scb->seqno) {
932 				if (--scb->pending == 0) {
933 					mtx_lock(&aio_job_mtx);
934 					scb->jobstate = JOBST_JOBQGLOBAL;
935 					TAILQ_REMOVE(&ki->kaio_syncqueue, scb, list);
936 					TAILQ_INSERT_TAIL(&aio_jobs, scb, list);
937 					aio_kick_nowait(userp);
938 					mtx_unlock(&aio_job_mtx);
939 				}
940 			}
941 		}
942 	}
943 	if (ki->kaio_flags & KAIO_WAKEUP) {
944 		ki->kaio_flags &= ~KAIO_WAKEUP;
945 		wakeup(&userp->p_aioinfo);
946 	}
947 }
948 
949 /*
950  * The AIO daemon, most of the actual work is done in aio_process,
951  * but the setup (and address space mgmt) is done in this routine.
952  */
953 static void
954 aio_daemon(void *_id)
955 {
956 	struct aiocblist *aiocbe;
957 	struct aiothreadlist *aiop;
958 	struct kaioinfo *ki;
959 	struct proc *curcp, *mycp, *userp;
960 	struct vmspace *myvm, *tmpvm;
961 	struct thread *td = curthread;
962 	int id = (intptr_t)_id;
963 
964 	/*
965 	 * Local copies of curproc (cp) and vmspace (myvm)
966 	 */
967 	mycp = td->td_proc;
968 	myvm = mycp->p_vmspace;
969 
970 	KASSERT(mycp->p_textvp == NULL, ("kthread has a textvp"));
971 
972 	/*
973 	 * Allocate and ready the aio control info.  There is one aiop structure
974 	 * per daemon.
975 	 */
976 	aiop = uma_zalloc(aiop_zone, M_WAITOK);
977 	aiop->aiothread = td;
978 	aiop->aiothreadflags = 0;
979 
980 	/* The daemon resides in its own pgrp. */
981 	setsid(td, NULL);
982 
983 	/*
984 	 * Wakeup parent process.  (Parent sleeps to keep from blasting away
985 	 * and creating too many daemons.)
986 	 */
987 	sema_post(&aio_newproc_sem);
988 
989 	mtx_lock(&aio_job_mtx);
990 	for (;;) {
991 		/*
992 		 * curcp is the current daemon process context.
993 		 * userp is the current user process context.
994 		 */
995 		curcp = mycp;
996 
997 		/*
998 		 * Take daemon off of free queue
999 		 */
1000 		if (aiop->aiothreadflags & AIOP_FREE) {
1001 			TAILQ_REMOVE(&aio_freeproc, aiop, list);
1002 			aiop->aiothreadflags &= ~AIOP_FREE;
1003 		}
1004 
1005 		/*
1006 		 * Check for jobs.
1007 		 */
1008 		while ((aiocbe = aio_selectjob(aiop)) != NULL) {
1009 			mtx_unlock(&aio_job_mtx);
1010 			userp = aiocbe->userproc;
1011 
1012 			/*
1013 			 * Connect to process address space for user program.
1014 			 */
1015 			if (userp != curcp) {
1016 				/*
1017 				 * Save the current address space that we are
1018 				 * connected to.
1019 				 */
1020 				tmpvm = mycp->p_vmspace;
1021 
1022 				/*
1023 				 * Point to the new user address space, and
1024 				 * refer to it.
1025 				 */
1026 				mycp->p_vmspace = userp->p_vmspace;
1027 				atomic_add_int(&mycp->p_vmspace->vm_refcnt, 1);
1028 
1029 				/* Activate the new mapping. */
1030 				pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1031 
1032 				/*
1033 				 * If the old address space wasn't the daemons
1034 				 * own address space, then we need to remove the
1035 				 * daemon's reference from the other process
1036 				 * that it was acting on behalf of.
1037 				 */
1038 				if (tmpvm != myvm) {
1039 					vmspace_free(tmpvm);
1040 				}
1041 				curcp = userp;
1042 			}
1043 
1044 			ki = userp->p_aioinfo;
1045 
1046 			/* Do the I/O function. */
1047 			aio_process(aiocbe);
1048 
1049 			mtx_lock(&aio_job_mtx);
1050 			/* Decrement the active job count. */
1051 			ki->kaio_active_count--;
1052 			mtx_unlock(&aio_job_mtx);
1053 
1054 			AIO_LOCK(ki);
1055 			TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
1056 			aio_bio_done_notify(userp, aiocbe, DONE_QUEUE);
1057 			AIO_UNLOCK(ki);
1058 
1059 			mtx_lock(&aio_job_mtx);
1060 		}
1061 
1062 		/*
1063 		 * Disconnect from user address space.
1064 		 */
1065 		if (curcp != mycp) {
1066 
1067 			mtx_unlock(&aio_job_mtx);
1068 
1069 			/* Get the user address space to disconnect from. */
1070 			tmpvm = mycp->p_vmspace;
1071 
1072 			/* Get original address space for daemon. */
1073 			mycp->p_vmspace = myvm;
1074 
1075 			/* Activate the daemon's address space. */
1076 			pmap_activate(FIRST_THREAD_IN_PROC(mycp));
1077 #ifdef DIAGNOSTIC
1078 			if (tmpvm == myvm) {
1079 				printf("AIOD: vmspace problem -- %d\n",
1080 				    mycp->p_pid);
1081 			}
1082 #endif
1083 			/* Remove our vmspace reference. */
1084 			vmspace_free(tmpvm);
1085 
1086 			curcp = mycp;
1087 
1088 			mtx_lock(&aio_job_mtx);
1089 			/*
1090 			 * We have to restart to avoid race, we only sleep if
1091 			 * no job can be selected, that should be
1092 			 * curcp == mycp.
1093 			 */
1094 			continue;
1095 		}
1096 
1097 		mtx_assert(&aio_job_mtx, MA_OWNED);
1098 
1099 		TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
1100 		aiop->aiothreadflags |= AIOP_FREE;
1101 
1102 		/*
1103 		 * If daemon is inactive for a long time, allow it to exit,
1104 		 * thereby freeing resources.
1105 		 */
1106 		if (msleep(aiop->aiothread, &aio_job_mtx, PRIBIO, "aiordy",
1107 		    aiod_lifetime)) {
1108 			if (TAILQ_EMPTY(&aio_jobs)) {
1109 				if ((aiop->aiothreadflags & AIOP_FREE) &&
1110 				    (num_aio_procs > target_aio_procs)) {
1111 					TAILQ_REMOVE(&aio_freeproc, aiop, list);
1112 					num_aio_procs--;
1113 					mtx_unlock(&aio_job_mtx);
1114 					uma_zfree(aiop_zone, aiop);
1115 					free_unr(aiod_unr, id);
1116 #ifdef DIAGNOSTIC
1117 					if (mycp->p_vmspace->vm_refcnt <= 1) {
1118 						printf("AIOD: bad vm refcnt for"
1119 						    " exiting daemon: %d\n",
1120 						    mycp->p_vmspace->vm_refcnt);
1121 					}
1122 #endif
1123 					kthread_exit(0);
1124 				}
1125 			}
1126 		}
1127 	}
1128 	mtx_unlock(&aio_job_mtx);
1129 	panic("shouldn't be here\n");
1130 }
1131 
1132 /*
1133  * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
1134  * AIO daemon modifies its environment itself.
1135  */
1136 static int
1137 aio_newproc(int *start)
1138 {
1139 	int error;
1140 	struct proc *p;
1141 	int id;
1142 
1143 	id = alloc_unr(aiod_unr);
1144 	error = kthread_create(aio_daemon, (void *)(intptr_t)id, &p,
1145 		RFNOWAIT, 0, "aiod%d", id);
1146 	if (error == 0) {
1147 		/*
1148 		 * Wait until daemon is started.
1149 		 */
1150 		sema_wait(&aio_newproc_sem);
1151 		mtx_lock(&aio_job_mtx);
1152 		num_aio_procs++;
1153 		if (start != NULL)
1154 			(*start)--;
1155 		mtx_unlock(&aio_job_mtx);
1156 	} else {
1157 		free_unr(aiod_unr, id);
1158 	}
1159 	return (error);
1160 }
1161 
1162 /*
1163  * Try the high-performance, low-overhead physio method for eligible
1164  * VCHR devices.  This method doesn't use an aio helper thread, and
1165  * thus has very low overhead.
1166  *
1167  * Assumes that the caller, aio_aqueue(), has incremented the file
1168  * structure's reference count, preventing its deallocation for the
1169  * duration of this call.
1170  */
1171 static int
1172 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
1173 {
1174 	struct aiocb *cb;
1175 	struct file *fp;
1176 	struct buf *bp;
1177 	struct vnode *vp;
1178 	struct kaioinfo *ki;
1179 	struct aioliojob *lj;
1180 	int error;
1181 
1182 	cb = &aiocbe->uaiocb;
1183 	fp = aiocbe->fd_file;
1184 
1185 	if (fp->f_type != DTYPE_VNODE)
1186 		return (-1);
1187 
1188 	vp = fp->f_vnode;
1189 
1190 	/*
1191 	 * If its not a disk, we don't want to return a positive error.
1192 	 * It causes the aio code to not fall through to try the thread
1193 	 * way when you're talking to a regular file.
1194 	 */
1195 	if (!vn_isdisk(vp, &error)) {
1196 		if (error == ENOTBLK)
1197 			return (-1);
1198 		else
1199 			return (error);
1200 	}
1201 
1202 	if (vp->v_bufobj.bo_bsize == 0)
1203 		return (-1);
1204 
1205  	if (cb->aio_nbytes % vp->v_bufobj.bo_bsize)
1206 		return (-1);
1207 
1208 	if (cb->aio_nbytes > vp->v_rdev->si_iosize_max)
1209 		return (-1);
1210 
1211 	if (cb->aio_nbytes >
1212 	    MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
1213 		return (-1);
1214 
1215 	ki = p->p_aioinfo;
1216 	if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
1217 		return (-1);
1218 
1219 	/* Create and build a buffer header for a transfer. */
1220 	bp = (struct buf *)getpbuf(NULL);
1221 	BUF_KERNPROC(bp);
1222 
1223 	AIO_LOCK(ki);
1224 	ki->kaio_count++;
1225 	ki->kaio_buffer_count++;
1226 	lj = aiocbe->lio;
1227 	if (lj)
1228 		lj->lioj_count++;
1229 	AIO_UNLOCK(ki);
1230 
1231 	/*
1232 	 * Get a copy of the kva from the physical buffer.
1233 	 */
1234 	error = 0;
1235 
1236 	bp->b_bcount = cb->aio_nbytes;
1237 	bp->b_bufsize = cb->aio_nbytes;
1238 	bp->b_iodone = aio_physwakeup;
1239 	bp->b_saveaddr = bp->b_data;
1240 	bp->b_data = (void *)(uintptr_t)cb->aio_buf;
1241 	bp->b_offset = cb->aio_offset;
1242 	bp->b_iooffset = cb->aio_offset;
1243 	bp->b_blkno = btodb(cb->aio_offset);
1244 	bp->b_iocmd = cb->aio_lio_opcode == LIO_WRITE ? BIO_WRITE : BIO_READ;
1245 
1246 	/*
1247 	 * Bring buffer into kernel space.
1248 	 */
1249 	if (vmapbuf(bp) < 0) {
1250 		error = EFAULT;
1251 		goto doerror;
1252 	}
1253 
1254 	AIO_LOCK(ki);
1255 	aiocbe->bp = bp;
1256 	bp->b_caller1 = (void *)aiocbe;
1257 	TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
1258 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1259 	aiocbe->jobstate = JOBST_JOBQBUF;
1260 	cb->_aiocb_private.status = cb->aio_nbytes;
1261 	AIO_UNLOCK(ki);
1262 
1263 	atomic_add_int(&num_queue_count, 1);
1264 	atomic_add_int(&num_buf_aio, 1);
1265 
1266 	bp->b_error = 0;
1267 
1268 	TASK_INIT(&aiocbe->biotask, 0, biohelper, aiocbe);
1269 
1270 	/* Perform transfer. */
1271 	dev_strategy(vp->v_rdev, bp);
1272 	return (0);
1273 
1274 doerror:
1275 	AIO_LOCK(ki);
1276 	ki->kaio_count--;
1277 	ki->kaio_buffer_count--;
1278 	if (lj)
1279 		lj->lioj_count--;
1280 	aiocbe->bp = NULL;
1281 	AIO_UNLOCK(ki);
1282 	relpbuf(bp, NULL);
1283 	return (error);
1284 }
1285 
1286 /*
1287  * Wake up aio requests that may be serviceable now.
1288  */
1289 static void
1290 aio_swake_cb(struct socket *so, struct sockbuf *sb)
1291 {
1292 	struct aiocblist *cb, *cbn;
1293 	int opcode;
1294 
1295 	if (sb == &so->so_snd)
1296 		opcode = LIO_WRITE;
1297 	else
1298 		opcode = LIO_READ;
1299 
1300 	SOCKBUF_LOCK(sb);
1301 	sb->sb_flags &= ~SB_AIO;
1302 	mtx_lock(&aio_job_mtx);
1303 	TAILQ_FOREACH_SAFE(cb, &so->so_aiojobq, list, cbn) {
1304 		if (opcode == cb->uaiocb.aio_lio_opcode) {
1305 			if (cb->jobstate != JOBST_JOBQSOCK)
1306 				panic("invalid queue value");
1307 			/* XXX
1308 			 * We don't have actual sockets backend yet,
1309 			 * so we simply move the requests to the generic
1310 			 * file I/O backend.
1311 			 */
1312 			TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1313 			TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1314 			aio_kick_nowait(cb->userproc);
1315 		}
1316 	}
1317 	mtx_unlock(&aio_job_mtx);
1318 	SOCKBUF_UNLOCK(sb);
1319 }
1320 
1321 /*
1322  * Queue a new AIO request.  Choosing either the threaded or direct physio VCHR
1323  * technique is done in this code.
1324  */
1325 int
1326 aio_aqueue(struct thread *td, struct aiocb *job, struct aioliojob *lj,
1327 	int type, int oldsigev)
1328 {
1329 	struct proc *p = td->td_proc;
1330 	struct file *fp;
1331 	struct socket *so;
1332 	struct aiocblist *aiocbe, *cb;
1333 	struct kaioinfo *ki;
1334 	struct kevent kev;
1335 	struct sockbuf *sb;
1336 	int opcode;
1337 	int error;
1338 	int fd, kqfd;
1339 	int jid;
1340 
1341 	if (p->p_aioinfo == NULL)
1342 		aio_init_aioinfo(p);
1343 
1344 	ki = p->p_aioinfo;
1345 
1346 	suword(&job->_aiocb_private.status, -1);
1347 	suword(&job->_aiocb_private.error, 0);
1348 	suword(&job->_aiocb_private.kernelinfo, -1);
1349 
1350 	if (num_queue_count >= max_queue_count ||
1351 	    ki->kaio_count >= ki->kaio_qallowed_count) {
1352 		suword(&job->_aiocb_private.error, EAGAIN);
1353 		return (EAGAIN);
1354 	}
1355 
1356 	aiocbe = uma_zalloc(aiocb_zone, M_WAITOK | M_ZERO);
1357 	aiocbe->inputcharge = 0;
1358 	aiocbe->outputcharge = 0;
1359 	knlist_init(&aiocbe->klist, AIO_MTX(ki), NULL, NULL, NULL);
1360 
1361 	if (oldsigev) {
1362 		bzero(&aiocbe->uaiocb, sizeof(struct aiocb));
1363 		error = copyin(job, &aiocbe->uaiocb, sizeof(struct oaiocb));
1364 		bcopy(&aiocbe->uaiocb.__spare__, &aiocbe->uaiocb.aio_sigevent,
1365 			sizeof(struct osigevent));
1366 	} else {
1367 		error = copyin(job, &aiocbe->uaiocb, sizeof(struct aiocb));
1368 	}
1369 	if (error) {
1370 		suword(&job->_aiocb_private.error, error);
1371 		uma_zfree(aiocb_zone, aiocbe);
1372 		return (error);
1373 	}
1374 
1375 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT &&
1376 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_SIGNAL &&
1377 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_THREAD_ID &&
1378 	    aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_NONE) {
1379 		suword(&job->_aiocb_private.error, EINVAL);
1380 		uma_zfree(aiocb_zone, aiocbe);
1381 		return (EINVAL);
1382 	}
1383 
1384 	if ((aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL ||
1385 	     aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_THREAD_ID) &&
1386 		!_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1387 		uma_zfree(aiocb_zone, aiocbe);
1388 		return (EINVAL);
1389 	}
1390 
1391 	ksiginfo_init(&aiocbe->ksi);
1392 
1393 	/* Save userspace address of the job info. */
1394 	aiocbe->uuaiocb = job;
1395 
1396 	/* Get the opcode. */
1397 	if (type != LIO_NOP)
1398 		aiocbe->uaiocb.aio_lio_opcode = type;
1399 	opcode = aiocbe->uaiocb.aio_lio_opcode;
1400 
1401 	/* Fetch the file object for the specified file descriptor. */
1402 	fd = aiocbe->uaiocb.aio_fildes;
1403 	switch (opcode) {
1404 	case LIO_WRITE:
1405 		error = fget_write(td, fd, &fp);
1406 		break;
1407 	case LIO_READ:
1408 		error = fget_read(td, fd, &fp);
1409 		break;
1410 	default:
1411 		error = fget(td, fd, &fp);
1412 	}
1413 	if (error) {
1414 		uma_zfree(aiocb_zone, aiocbe);
1415 		suword(&job->_aiocb_private.error, error);
1416 		return (error);
1417 	}
1418 
1419 	if (opcode == LIO_SYNC && fp->f_vnode == NULL) {
1420 		error = EINVAL;
1421 		goto aqueue_fail;
1422 	}
1423 
1424 	if (opcode != LIO_SYNC && aiocbe->uaiocb.aio_offset == -1LL) {
1425 		error = EINVAL;
1426 		goto aqueue_fail;
1427 	}
1428 
1429 	aiocbe->fd_file = fp;
1430 
1431 	mtx_lock(&aio_job_mtx);
1432 	jid = jobrefid++;
1433 	aiocbe->seqno = jobseqno++;
1434 	mtx_unlock(&aio_job_mtx);
1435 	error = suword(&job->_aiocb_private.kernelinfo, jid);
1436 	if (error) {
1437 		error = EINVAL;
1438 		goto aqueue_fail;
1439 	}
1440 	aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jid;
1441 
1442 	if (opcode == LIO_NOP) {
1443 		fdrop(fp, td);
1444 		uma_zfree(aiocb_zone, aiocbe);
1445 		return (0);
1446 	}
1447 	if ((opcode != LIO_READ) && (opcode != LIO_WRITE) &&
1448 	    (opcode != LIO_SYNC)) {
1449 		error = EINVAL;
1450 		goto aqueue_fail;
1451 	}
1452 
1453 	if (aiocbe->uaiocb.aio_sigevent.sigev_notify != SIGEV_KEVENT)
1454 		goto no_kqueue;
1455 	kqfd = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1456 	kev.ident = (uintptr_t)aiocbe->uuaiocb;
1457 	kev.filter = EVFILT_AIO;
1458 	kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1459 	kev.data = (intptr_t)aiocbe;
1460 	kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1461 	error = kqfd_register(kqfd, &kev, td, 1);
1462 aqueue_fail:
1463 	if (error) {
1464 		fdrop(fp, td);
1465 		uma_zfree(aiocb_zone, aiocbe);
1466 		suword(&job->_aiocb_private.error, error);
1467 		goto done;
1468 	}
1469 no_kqueue:
1470 
1471 	suword(&job->_aiocb_private.error, EINPROGRESS);
1472 	aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1473 	aiocbe->userproc = p;
1474 	aiocbe->cred = crhold(td->td_ucred);
1475 	aiocbe->jobflags = 0;
1476 	aiocbe->lio = lj;
1477 
1478 	if (opcode == LIO_SYNC)
1479 		goto queueit;
1480 
1481 	if (fp->f_type == DTYPE_SOCKET) {
1482 		/*
1483 		 * Alternate queueing for socket ops: Reach down into the
1484 		 * descriptor to get the socket data.  Then check to see if the
1485 		 * socket is ready to be read or written (based on the requested
1486 		 * operation).
1487 		 *
1488 		 * If it is not ready for io, then queue the aiocbe on the
1489 		 * socket, and set the flags so we get a call when sbnotify()
1490 		 * happens.
1491 		 *
1492 		 * Note if opcode is neither LIO_WRITE nor LIO_READ we lock
1493 		 * and unlock the snd sockbuf for no reason.
1494 		 */
1495 		so = fp->f_data;
1496 		sb = (opcode == LIO_READ) ? &so->so_rcv : &so->so_snd;
1497 		SOCKBUF_LOCK(sb);
1498 		if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1499 		    LIO_WRITE) && (!sowriteable(so)))) {
1500 			sb->sb_flags |= SB_AIO;
1501 
1502 			mtx_lock(&aio_job_mtx);
1503 			TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1504 			mtx_unlock(&aio_job_mtx);
1505 
1506 			AIO_LOCK(ki);
1507 			TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1508 			TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1509 			aiocbe->jobstate = JOBST_JOBQSOCK;
1510 			ki->kaio_count++;
1511 			if (lj)
1512 				lj->lioj_count++;
1513 			AIO_UNLOCK(ki);
1514 			SOCKBUF_UNLOCK(sb);
1515 			atomic_add_int(&num_queue_count, 1);
1516 			error = 0;
1517 			goto done;
1518 		}
1519 		SOCKBUF_UNLOCK(sb);
1520 	}
1521 
1522 	if ((error = aio_qphysio(p, aiocbe)) == 0)
1523 		goto done;
1524 #if 0
1525 	if (error > 0) {
1526 		aiocbe->uaiocb._aiocb_private.error = error;
1527 		suword(&job->_aiocb_private.error, error);
1528 		goto done;
1529 	}
1530 #endif
1531 queueit:
1532 	/* No buffer for daemon I/O. */
1533 	aiocbe->bp = NULL;
1534 	atomic_add_int(&num_queue_count, 1);
1535 
1536 	AIO_LOCK(ki);
1537 	ki->kaio_count++;
1538 	if (lj)
1539 		lj->lioj_count++;
1540 	TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1541 	TAILQ_INSERT_TAIL(&ki->kaio_all, aiocbe, allist);
1542 	if (opcode == LIO_SYNC) {
1543 		TAILQ_FOREACH(cb, &ki->kaio_jobqueue, plist) {
1544 			if (cb->fd_file == aiocbe->fd_file &&
1545 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1546 			    cb->seqno < aiocbe->seqno) {
1547 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1548 				aiocbe->pending++;
1549 			}
1550 		}
1551 		TAILQ_FOREACH(cb, &ki->kaio_bufqueue, plist) {
1552 			if (cb->fd_file == aiocbe->fd_file &&
1553 			    cb->uaiocb.aio_lio_opcode != LIO_SYNC &&
1554 			    cb->seqno < aiocbe->seqno) {
1555 				cb->jobflags |= AIOCBLIST_CHECKSYNC;
1556 				aiocbe->pending++;
1557 			}
1558 		}
1559 		if (aiocbe->pending != 0) {
1560 			TAILQ_INSERT_TAIL(&ki->kaio_syncqueue, aiocbe, list);
1561 			aiocbe->jobstate = JOBST_JOBQSYNC;
1562 			AIO_UNLOCK(ki);
1563 			goto done;
1564 		}
1565 	}
1566 	mtx_lock(&aio_job_mtx);
1567 	TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1568 	aiocbe->jobstate = JOBST_JOBQGLOBAL;
1569 	aio_kick_nowait(p);
1570 	mtx_unlock(&aio_job_mtx);
1571 	AIO_UNLOCK(ki);
1572 	error = 0;
1573 done:
1574 	return (error);
1575 }
1576 
1577 static void
1578 aio_kick_nowait(struct proc *userp)
1579 {
1580 	struct kaioinfo *ki = userp->p_aioinfo;
1581 	struct aiothreadlist *aiop;
1582 
1583 	mtx_assert(&aio_job_mtx, MA_OWNED);
1584 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1585 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1586 		aiop->aiothreadflags &= ~AIOP_FREE;
1587 		wakeup(aiop->aiothread);
1588 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1589 	    ((ki->kaio_active_count + num_aio_resv_start) <
1590 	    ki->kaio_maxactive_count)) {
1591 		taskqueue_enqueue(taskqueue_aiod_bio, &ki->kaio_task);
1592 	}
1593 }
1594 
1595 static int
1596 aio_kick(struct proc *userp)
1597 {
1598 	struct kaioinfo *ki = userp->p_aioinfo;
1599 	struct aiothreadlist *aiop;
1600 	int error, ret = 0;
1601 
1602 	mtx_assert(&aio_job_mtx, MA_OWNED);
1603 retryproc:
1604 	if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1605 		TAILQ_REMOVE(&aio_freeproc, aiop, list);
1606 		aiop->aiothreadflags &= ~AIOP_FREE;
1607 		wakeup(aiop->aiothread);
1608 	} else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1609 	    ((ki->kaio_active_count + num_aio_resv_start) <
1610 	    ki->kaio_maxactive_count)) {
1611 		num_aio_resv_start++;
1612 		mtx_unlock(&aio_job_mtx);
1613 		error = aio_newproc(&num_aio_resv_start);
1614 		mtx_lock(&aio_job_mtx);
1615 		if (error) {
1616 			num_aio_resv_start--;
1617 			goto retryproc;
1618 		}
1619 	} else {
1620 		ret = -1;
1621 	}
1622 	return (ret);
1623 }
1624 
1625 static void
1626 aio_kick_helper(void *context, int pending)
1627 {
1628 	struct proc *userp = context;
1629 
1630 	mtx_lock(&aio_job_mtx);
1631 	while (--pending >= 0) {
1632 		if (aio_kick(userp))
1633 			break;
1634 	}
1635 	mtx_unlock(&aio_job_mtx);
1636 }
1637 
1638 /*
1639  * Support the aio_return system call, as a side-effect, kernel resources are
1640  * released.
1641  */
1642 int
1643 aio_return(struct thread *td, struct aio_return_args *uap)
1644 {
1645 	struct proc *p = td->td_proc;
1646 	struct aiocblist *cb;
1647 	struct aiocb *uaiocb;
1648 	struct kaioinfo *ki;
1649 	int status, error;
1650 
1651 	ki = p->p_aioinfo;
1652 	if (ki == NULL)
1653 		return (EINVAL);
1654 	uaiocb = uap->aiocbp;
1655 	AIO_LOCK(ki);
1656 	TAILQ_FOREACH(cb, &ki->kaio_done, plist) {
1657 		if (cb->uuaiocb == uaiocb)
1658 			break;
1659 	}
1660 	if (cb != NULL) {
1661 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
1662 		status = cb->uaiocb._aiocb_private.status;
1663 		error = cb->uaiocb._aiocb_private.error;
1664 		td->td_retval[0] = status;
1665 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1666 			p->p_stats->p_ru.ru_oublock +=
1667 			    cb->outputcharge;
1668 			cb->outputcharge = 0;
1669 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1670 			p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1671 			cb->inputcharge = 0;
1672 		}
1673 		aio_free_entry(cb);
1674 		AIO_UNLOCK(ki);
1675 		suword(&uaiocb->_aiocb_private.error, error);
1676 		suword(&uaiocb->_aiocb_private.status, status);
1677 	} else {
1678 		error = EINVAL;
1679 		AIO_UNLOCK(ki);
1680 	}
1681 	return (error);
1682 }
1683 
1684 /*
1685  * Allow a process to wakeup when any of the I/O requests are completed.
1686  */
1687 int
1688 aio_suspend(struct thread *td, struct aio_suspend_args *uap)
1689 {
1690 	struct proc *p = td->td_proc;
1691 	struct timeval atv;
1692 	struct timespec ts;
1693 	struct aiocb *const *cbptr, *cbp;
1694 	struct kaioinfo *ki;
1695 	struct aiocblist *cb, *cbfirst;
1696 	struct aiocb **ujoblist;
1697 	int njoblist;
1698 	int error;
1699 	int timo;
1700 	int i;
1701 
1702 	if (uap->nent < 0 || uap->nent > AIO_LISTIO_MAX)
1703 		return (EINVAL);
1704 
1705 	timo = 0;
1706 	if (uap->timeout) {
1707 		/* Get timespec struct. */
1708 		if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1709 			return (error);
1710 
1711 		if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1712 			return (EINVAL);
1713 
1714 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
1715 		if (itimerfix(&atv))
1716 			return (EINVAL);
1717 		timo = tvtohz(&atv);
1718 	}
1719 
1720 	ki = p->p_aioinfo;
1721 	if (ki == NULL)
1722 		return (EAGAIN);
1723 
1724 	njoblist = 0;
1725 	ujoblist = uma_zalloc(aiol_zone, M_WAITOK);
1726 	cbptr = uap->aiocbp;
1727 
1728 	for (i = 0; i < uap->nent; i++) {
1729 		cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1730 		if (cbp == 0)
1731 			continue;
1732 		ujoblist[njoblist] = cbp;
1733 		njoblist++;
1734 	}
1735 
1736 	if (njoblist == 0) {
1737 		uma_zfree(aiol_zone, ujoblist);
1738 		return (0);
1739 	}
1740 
1741 	AIO_LOCK(ki);
1742 	for (;;) {
1743 		cbfirst = NULL;
1744 		error = 0;
1745 		TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1746 			for (i = 0; i < njoblist; i++) {
1747 				if (cb->uuaiocb == ujoblist[i]) {
1748 					if (cbfirst == NULL)
1749 						cbfirst = cb;
1750 					if (cb->jobstate == JOBST_JOBFINISHED)
1751 						goto RETURN;
1752 				}
1753 			}
1754 		}
1755 		/* All tasks were finished. */
1756 		if (cbfirst == NULL)
1757 			break;
1758 
1759 		ki->kaio_flags |= KAIO_WAKEUP;
1760 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
1761 		    "aiospn", timo);
1762 		if (error == ERESTART)
1763 			error = EINTR;
1764 		if (error)
1765 			break;
1766 	}
1767 RETURN:
1768 	AIO_UNLOCK(ki);
1769 	uma_zfree(aiol_zone, ujoblist);
1770 	return (error);
1771 }
1772 
1773 /*
1774  * aio_cancel cancels any non-physio aio operations not currently in
1775  * progress.
1776  */
1777 int
1778 aio_cancel(struct thread *td, struct aio_cancel_args *uap)
1779 {
1780 	struct proc *p = td->td_proc;
1781 	struct kaioinfo *ki;
1782 	struct aiocblist *cbe, *cbn;
1783 	struct file *fp;
1784 	struct socket *so;
1785 	int error;
1786 	int remove;
1787 	int cancelled = 0;
1788 	int notcancelled = 0;
1789 	struct vnode *vp;
1790 
1791 	/* Lookup file object. */
1792 	error = fget(td, uap->fd, &fp);
1793 	if (error)
1794 		return (error);
1795 
1796 	ki = p->p_aioinfo;
1797 	if (ki == NULL)
1798 		goto done;
1799 
1800 	if (fp->f_type == DTYPE_VNODE) {
1801 		vp = fp->f_vnode;
1802 		if (vn_isdisk(vp, &error)) {
1803 			fdrop(fp, td);
1804 			td->td_retval[0] = AIO_NOTCANCELED;
1805 			return (0);
1806 		}
1807 	}
1808 
1809 	AIO_LOCK(ki);
1810 	TAILQ_FOREACH_SAFE(cbe, &ki->kaio_jobqueue, plist, cbn) {
1811 		if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1812 		    ((uap->aiocbp == NULL) ||
1813 		     (uap->aiocbp == cbe->uuaiocb))) {
1814 			remove = 0;
1815 
1816 			mtx_lock(&aio_job_mtx);
1817 			if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1818 				TAILQ_REMOVE(&aio_jobs, cbe, list);
1819 				remove = 1;
1820 			} else if (cbe->jobstate == JOBST_JOBQSOCK) {
1821 				MPASS(fp->f_type == DTYPE_SOCKET);
1822 				so = fp->f_data;
1823 				TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1824 				remove = 1;
1825 			} else if (cbe->jobstate == JOBST_JOBQSYNC) {
1826 				TAILQ_REMOVE(&ki->kaio_syncqueue, cbe, list);
1827 				remove = 1;
1828 			}
1829 			mtx_unlock(&aio_job_mtx);
1830 
1831 			if (remove) {
1832 				TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1833 				cbe->uaiocb._aiocb_private.status = -1;
1834 				cbe->uaiocb._aiocb_private.error = ECANCELED;
1835 				aio_bio_done_notify(p, cbe, DONE_QUEUE);
1836 				cancelled++;
1837 			} else {
1838 				notcancelled++;
1839 			}
1840 			if (uap->aiocbp != NULL)
1841 				break;
1842 		}
1843 	}
1844 	AIO_UNLOCK(ki);
1845 
1846 done:
1847 	fdrop(fp, td);
1848 
1849 	if (uap->aiocbp != NULL) {
1850 		if (cancelled) {
1851 			td->td_retval[0] = AIO_CANCELED;
1852 			return (0);
1853 		}
1854 	}
1855 
1856 	if (notcancelled) {
1857 		td->td_retval[0] = AIO_NOTCANCELED;
1858 		return (0);
1859 	}
1860 
1861 	if (cancelled) {
1862 		td->td_retval[0] = AIO_CANCELED;
1863 		return (0);
1864 	}
1865 
1866 	td->td_retval[0] = AIO_ALLDONE;
1867 
1868 	return (0);
1869 }
1870 
1871 /*
1872  * aio_error is implemented in the kernel level for compatibility purposes
1873  * only.  For a user mode async implementation, it would be best to do it in
1874  * a userland subroutine.
1875  */
1876 int
1877 aio_error(struct thread *td, struct aio_error_args *uap)
1878 {
1879 	struct proc *p = td->td_proc;
1880 	struct aiocblist *cb;
1881 	struct kaioinfo *ki;
1882 	int status;
1883 
1884 	ki = p->p_aioinfo;
1885 	if (ki == NULL) {
1886 		td->td_retval[0] = EINVAL;
1887 		return (0);
1888 	}
1889 
1890 	AIO_LOCK(ki);
1891 	TAILQ_FOREACH(cb, &ki->kaio_all, allist) {
1892 		if (cb->uuaiocb == uap->aiocbp) {
1893 			if (cb->jobstate == JOBST_JOBFINISHED)
1894 				td->td_retval[0] =
1895 					cb->uaiocb._aiocb_private.error;
1896 			else
1897 				td->td_retval[0] = EINPROGRESS;
1898 			AIO_UNLOCK(ki);
1899 			return (0);
1900 		}
1901 	}
1902 	AIO_UNLOCK(ki);
1903 
1904 	/*
1905 	 * Hack for failure of aio_aqueue.
1906 	 */
1907 	status = fuword(&uap->aiocbp->_aiocb_private.status);
1908 	if (status == -1) {
1909 		td->td_retval[0] = fuword(&uap->aiocbp->_aiocb_private.error);
1910 		return (0);
1911 	}
1912 
1913 	td->td_retval[0] = EINVAL;
1914 	return (0);
1915 }
1916 
1917 /* syscall - asynchronous read from a file (REALTIME) */
1918 int
1919 oaio_read(struct thread *td, struct oaio_read_args *uap)
1920 {
1921 
1922 	return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_READ, 1);
1923 }
1924 
1925 int
1926 aio_read(struct thread *td, struct aio_read_args *uap)
1927 {
1928 
1929 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_READ, 0);
1930 }
1931 
1932 /* syscall - asynchronous write to a file (REALTIME) */
1933 int
1934 oaio_write(struct thread *td, struct oaio_write_args *uap)
1935 {
1936 
1937 	return aio_aqueue(td, (struct aiocb *)uap->aiocbp, NULL, LIO_WRITE, 1);
1938 }
1939 
1940 int
1941 aio_write(struct thread *td, struct aio_write_args *uap)
1942 {
1943 
1944 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_WRITE, 0);
1945 }
1946 
1947 /* syscall - list directed I/O (REALTIME) */
1948 int
1949 olio_listio(struct thread *td, struct olio_listio_args *uap)
1950 {
1951 	return do_lio_listio(td, (struct lio_listio_args *)uap, 1);
1952 }
1953 
1954 /* syscall - list directed I/O (REALTIME) */
1955 int
1956 lio_listio(struct thread *td, struct lio_listio_args *uap)
1957 {
1958 	return do_lio_listio(td, uap, 0);
1959 }
1960 
1961 static int
1962 do_lio_listio(struct thread *td, struct lio_listio_args *uap, int oldsigev)
1963 {
1964 	struct proc *p = td->td_proc;
1965 	struct aiocb *iocb, * const *cbptr;
1966 	struct kaioinfo *ki;
1967 	struct aioliojob *lj;
1968 	struct kevent kev;
1969 	int nent;
1970 	int error;
1971 	int nerror;
1972 	int i;
1973 
1974 	if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1975 		return (EINVAL);
1976 
1977 	nent = uap->nent;
1978 	if (nent < 0 || nent > AIO_LISTIO_MAX)
1979 		return (EINVAL);
1980 
1981 	if (p->p_aioinfo == NULL)
1982 		aio_init_aioinfo(p);
1983 
1984 	ki = p->p_aioinfo;
1985 
1986 	lj = uma_zalloc(aiolio_zone, M_WAITOK);
1987 	lj->lioj_flags = 0;
1988 	lj->lioj_count = 0;
1989 	lj->lioj_finished_count = 0;
1990 	knlist_init(&lj->klist, AIO_MTX(ki), NULL, NULL, NULL);
1991 	ksiginfo_init(&lj->lioj_ksi);
1992 
1993 	/*
1994 	 * Setup signal.
1995 	 */
1996 	if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1997 		bzero(&lj->lioj_signal, sizeof(&lj->lioj_signal));
1998 		error = copyin(uap->sig, &lj->lioj_signal,
1999 				oldsigev ? sizeof(struct osigevent) :
2000 					   sizeof(struct sigevent));
2001 		if (error) {
2002 			uma_zfree(aiolio_zone, lj);
2003 			return (error);
2004 		}
2005 
2006 		if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2007 			/* Assume only new style KEVENT */
2008 			kev.filter = EVFILT_LIO;
2009 			kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
2010 			kev.ident = (uintptr_t)uap->acb_list; /* something unique */
2011 			kev.data = (intptr_t)lj;
2012 			/* pass user defined sigval data */
2013 			kev.udata = lj->lioj_signal.sigev_value.sival_ptr;
2014 			error = kqfd_register(
2015 			    lj->lioj_signal.sigev_notify_kqueue, &kev, td, 1);
2016 			if (error) {
2017 				uma_zfree(aiolio_zone, lj);
2018 				return (error);
2019 			}
2020 		} else if (lj->lioj_signal.sigev_notify == SIGEV_NONE) {
2021 			;
2022 		} else if (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2023 			   lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID) {
2024 				if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
2025 					uma_zfree(aiolio_zone, lj);
2026 					return EINVAL;
2027 				}
2028 				lj->lioj_flags |= LIOJ_SIGNAL;
2029 		} else {
2030 			uma_zfree(aiolio_zone, lj);
2031 			return EINVAL;
2032 		}
2033 	}
2034 
2035 	AIO_LOCK(ki);
2036 	TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
2037 	/*
2038 	 * Add extra aiocb count to avoid the lio to be freed
2039 	 * by other threads doing aio_waitcomplete or aio_return,
2040 	 * and prevent event from being sent until we have queued
2041 	 * all tasks.
2042 	 */
2043 	lj->lioj_count = 1;
2044 	AIO_UNLOCK(ki);
2045 
2046 	/*
2047 	 * Get pointers to the list of I/O requests.
2048 	 */
2049 	nerror = 0;
2050 	cbptr = uap->acb_list;
2051 	for (i = 0; i < uap->nent; i++) {
2052 		iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
2053 		if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
2054 			error = aio_aqueue(td, iocb, lj, LIO_NOP, oldsigev);
2055 			if (error != 0)
2056 				nerror++;
2057 		}
2058 	}
2059 
2060 	error = 0;
2061 	AIO_LOCK(ki);
2062 	if (uap->mode == LIO_WAIT) {
2063 		while (lj->lioj_count - 1 != lj->lioj_finished_count) {
2064 			ki->kaio_flags |= KAIO_WAKEUP;
2065 			error = msleep(&p->p_aioinfo, AIO_MTX(ki),
2066 			    PRIBIO | PCATCH, "aiospn", 0);
2067 			if (error == ERESTART)
2068 				error = EINTR;
2069 			if (error)
2070 				break;
2071 		}
2072 	} else {
2073 		if (lj->lioj_count - 1 == lj->lioj_finished_count) {
2074 			if (lj->lioj_signal.sigev_notify == SIGEV_KEVENT) {
2075 				lj->lioj_flags |= LIOJ_KEVENT_POSTED;
2076 				KNOTE_LOCKED(&lj->klist, 1);
2077 			}
2078 			if ((lj->lioj_flags & (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED))
2079 			    == LIOJ_SIGNAL
2080 			    && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL ||
2081 			    lj->lioj_signal.sigev_notify == SIGEV_THREAD_ID)) {
2082 				aio_sendsig(p, &lj->lioj_signal,
2083 					    &lj->lioj_ksi);
2084 				lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2085 			}
2086 		}
2087 	}
2088 	lj->lioj_count--;
2089 	if (lj->lioj_count == 0) {
2090 		TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
2091 		knlist_delete(&lj->klist, curthread, 1);
2092 		PROC_LOCK(p);
2093 		sigqueue_take(&lj->lioj_ksi);
2094 		PROC_UNLOCK(p);
2095 		AIO_UNLOCK(ki);
2096 		uma_zfree(aiolio_zone, lj);
2097 	} else
2098 		AIO_UNLOCK(ki);
2099 
2100 	if (nerror)
2101 		return (EIO);
2102 	return (error);
2103 }
2104 
2105 /*
2106  * Called from interrupt thread for physio, we should return as fast
2107  * as possible, so we schedule a biohelper task.
2108  */
2109 static void
2110 aio_physwakeup(struct buf *bp)
2111 {
2112 	struct aiocblist *aiocbe;
2113 
2114 	aiocbe = (struct aiocblist *)bp->b_caller1;
2115 	taskqueue_enqueue(taskqueue_aiod_bio, &aiocbe->biotask);
2116 }
2117 
2118 /*
2119  * Task routine to perform heavy tasks, process wakeup, and signals.
2120  */
2121 static void
2122 biohelper(void *context, int pending)
2123 {
2124 	struct aiocblist *aiocbe = context;
2125 	struct buf *bp;
2126 	struct proc *userp;
2127 	struct kaioinfo *ki;
2128 	int nblks;
2129 
2130 	bp = aiocbe->bp;
2131 	userp = aiocbe->userproc;
2132 	ki = userp->p_aioinfo;
2133 	AIO_LOCK(ki);
2134 	aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2135 	aiocbe->uaiocb._aiocb_private.error = 0;
2136 	if (bp->b_ioflags & BIO_ERROR)
2137 		aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2138 	nblks = btodb(aiocbe->uaiocb.aio_nbytes);
2139 	if (aiocbe->uaiocb.aio_lio_opcode == LIO_WRITE)
2140 		aiocbe->outputcharge += nblks;
2141 	else
2142 		aiocbe->inputcharge += nblks;
2143 	aiocbe->bp = NULL;
2144 	TAILQ_REMOVE(&userp->p_aioinfo->kaio_bufqueue, aiocbe, plist);
2145 	ki->kaio_buffer_count--;
2146 	aio_bio_done_notify(userp, aiocbe, DONE_BUF);
2147 	AIO_UNLOCK(ki);
2148 
2149 	/* Release mapping into kernel space. */
2150 	vunmapbuf(bp);
2151 	relpbuf(bp, NULL);
2152 	atomic_subtract_int(&num_buf_aio, 1);
2153 }
2154 
2155 /* syscall - wait for the next completion of an aio request */
2156 int
2157 aio_waitcomplete(struct thread *td, struct aio_waitcomplete_args *uap)
2158 {
2159 	struct proc *p = td->td_proc;
2160 	struct timeval atv;
2161 	struct timespec ts;
2162 	struct kaioinfo *ki;
2163 	struct aiocblist *cb;
2164 	struct aiocb *uuaiocb;
2165 	int error, status, timo;
2166 
2167 	suword(uap->aiocbp, (long)NULL);
2168 
2169 	timo = 0;
2170 	if (uap->timeout) {
2171 		/* Get timespec struct. */
2172 		error = copyin(uap->timeout, &ts, sizeof(ts));
2173 		if (error)
2174 			return (error);
2175 
2176 		if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2177 			return (EINVAL);
2178 
2179 		TIMESPEC_TO_TIMEVAL(&atv, &ts);
2180 		if (itimerfix(&atv))
2181 			return (EINVAL);
2182 		timo = tvtohz(&atv);
2183 	}
2184 
2185 	if (p->p_aioinfo == NULL)
2186 		aio_init_aioinfo(p);
2187 	ki = p->p_aioinfo;
2188 
2189 	error = 0;
2190 	cb = NULL;
2191 	AIO_LOCK(ki);
2192 	while ((cb = TAILQ_FIRST(&ki->kaio_done)) == NULL) {
2193 		ki->kaio_flags |= KAIO_WAKEUP;
2194 		error = msleep(&p->p_aioinfo, AIO_MTX(ki), PRIBIO | PCATCH,
2195 		    "aiowc", timo);
2196 		if (timo && error == ERESTART)
2197 			error = EINTR;
2198 		if (error)
2199 			break;
2200 	}
2201 
2202 	if (cb != NULL) {
2203 		MPASS(cb->jobstate == JOBST_JOBFINISHED);
2204 		uuaiocb = cb->uuaiocb;
2205 		status = cb->uaiocb._aiocb_private.status;
2206 		error = cb->uaiocb._aiocb_private.error;
2207 		td->td_retval[0] = status;
2208 		if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2209 			p->p_stats->p_ru.ru_oublock += cb->outputcharge;
2210 			cb->outputcharge = 0;
2211 		} else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2212 			p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2213 			cb->inputcharge = 0;
2214 		}
2215 		aio_free_entry(cb);
2216 		AIO_UNLOCK(ki);
2217 		suword(uap->aiocbp, (long)uuaiocb);
2218 		suword(&uuaiocb->_aiocb_private.error, error);
2219 		suword(&uuaiocb->_aiocb_private.status, status);
2220 	} else
2221 		AIO_UNLOCK(ki);
2222 
2223 	return (error);
2224 }
2225 
2226 int
2227 aio_fsync(struct thread *td, struct aio_fsync_args *uap)
2228 {
2229 	struct proc *p = td->td_proc;
2230 	struct kaioinfo *ki;
2231 
2232 	if (uap->op != O_SYNC) /* XXX lack of O_DSYNC */
2233 		return (EINVAL);
2234 	ki = p->p_aioinfo;
2235 	if (ki == NULL)
2236 		aio_init_aioinfo(p);
2237 	return aio_aqueue(td, uap->aiocbp, NULL, LIO_SYNC, 0);
2238 }
2239 
2240 /* kqueue attach function */
2241 static int
2242 filt_aioattach(struct knote *kn)
2243 {
2244 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2245 
2246 	/*
2247 	 * The aiocbe pointer must be validated before using it, so
2248 	 * registration is restricted to the kernel; the user cannot
2249 	 * set EV_FLAG1.
2250 	 */
2251 	if ((kn->kn_flags & EV_FLAG1) == 0)
2252 		return (EPERM);
2253 	kn->kn_flags &= ~EV_FLAG1;
2254 
2255 	knlist_add(&aiocbe->klist, kn, 0);
2256 
2257 	return (0);
2258 }
2259 
2260 /* kqueue detach function */
2261 static void
2262 filt_aiodetach(struct knote *kn)
2263 {
2264 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2265 
2266 	if (!knlist_empty(&aiocbe->klist))
2267 		knlist_remove(&aiocbe->klist, kn, 0);
2268 }
2269 
2270 /* kqueue filter function */
2271 /*ARGSUSED*/
2272 static int
2273 filt_aio(struct knote *kn, long hint)
2274 {
2275 	struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2276 
2277 	kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2278 	if (aiocbe->jobstate != JOBST_JOBFINISHED)
2279 		return (0);
2280 	kn->kn_flags |= EV_EOF;
2281 	return (1);
2282 }
2283 
2284 /* kqueue attach function */
2285 static int
2286 filt_lioattach(struct knote *kn)
2287 {
2288 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2289 
2290 	/*
2291 	 * The aioliojob pointer must be validated before using it, so
2292 	 * registration is restricted to the kernel; the user cannot
2293 	 * set EV_FLAG1.
2294 	 */
2295 	if ((kn->kn_flags & EV_FLAG1) == 0)
2296 		return (EPERM);
2297 	kn->kn_flags &= ~EV_FLAG1;
2298 
2299 	knlist_add(&lj->klist, kn, 0);
2300 
2301 	return (0);
2302 }
2303 
2304 /* kqueue detach function */
2305 static void
2306 filt_liodetach(struct knote *kn)
2307 {
2308 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2309 
2310 	if (!knlist_empty(&lj->klist))
2311 		knlist_remove(&lj->klist, kn, 0);
2312 }
2313 
2314 /* kqueue filter function */
2315 /*ARGSUSED*/
2316 static int
2317 filt_lio(struct knote *kn, long hint)
2318 {
2319 	struct aioliojob * lj = (struct aioliojob *)kn->kn_sdata;
2320 
2321 	return (lj->lioj_flags & LIOJ_KEVENT_POSTED);
2322 }
2323