2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
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.
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.
16 * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.42 2007/07/20 17:21:52 dillon Exp $
21 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
24 #include <sys/param.h>
25 #include <sys/systm.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
33 #include <sys/unistd.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
42 #include <sys/event.h>
45 #include <vm/vm_extern.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
50 #include <sys/file2.h>
52 #include <sys/sysref2.h>
53 #include <sys/thread2.h>
55 #include <machine/limits.h>
56 #include "opt_vfs_aio.h"
61 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
66 #define JOBST_NULL 0x0
67 #define JOBST_JOBQGLOBAL 0x2
68 #define JOBST_JOBRUNNING 0x3
69 #define JOBST_JOBFINISHED 0x4
70 #define JOBST_JOBQBUF 0x5
71 #define JOBST_JOBBFINISHED 0x6
73 #ifndef MAX_AIO_PER_PROC
74 #define MAX_AIO_PER_PROC 32
77 #ifndef MAX_AIO_QUEUE_PER_PROC
78 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
82 #define MAX_AIO_PROCS 32
86 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
89 #ifndef TARGET_AIO_PROCS
90 #define TARGET_AIO_PROCS 4
94 #define MAX_BUF_AIO 16
97 #ifndef AIOD_TIMEOUT_DEFAULT
98 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
101 #ifndef AIOD_LIFETIME_DEFAULT
102 #define AIOD_LIFETIME_DEFAULT (30 * hz)
105 SYSCTL_NODE(_vfs
, OID_AUTO
, aio
, CTLFLAG_RW
, 0, "Async IO management");
107 static int max_aio_procs
= MAX_AIO_PROCS
;
108 SYSCTL_INT(_vfs_aio
, OID_AUTO
, max_aio_procs
,
109 CTLFLAG_RW
, &max_aio_procs
, 0,
110 "Maximum number of kernel threads to use for handling async IO");
112 static int num_aio_procs
= 0;
113 SYSCTL_INT(_vfs_aio
, OID_AUTO
, num_aio_procs
,
114 CTLFLAG_RD
, &num_aio_procs
, 0,
115 "Number of presently active kernel threads for async IO");
118 * The code will adjust the actual number of AIO processes towards this
119 * number when it gets a chance.
121 static int target_aio_procs
= TARGET_AIO_PROCS
;
122 SYSCTL_INT(_vfs_aio
, OID_AUTO
, target_aio_procs
, CTLFLAG_RW
, &target_aio_procs
,
123 0, "Preferred number of ready kernel threads for async IO");
125 static int max_queue_count
= MAX_AIO_QUEUE
;
126 SYSCTL_INT(_vfs_aio
, OID_AUTO
, max_aio_queue
, CTLFLAG_RW
, &max_queue_count
, 0,
127 "Maximum number of aio requests to queue, globally");
129 static int num_queue_count
= 0;
130 SYSCTL_INT(_vfs_aio
, OID_AUTO
, num_queue_count
, CTLFLAG_RD
, &num_queue_count
, 0,
131 "Number of queued aio requests");
133 static int num_buf_aio
= 0;
134 SYSCTL_INT(_vfs_aio
, OID_AUTO
, num_buf_aio
, CTLFLAG_RD
, &num_buf_aio
, 0,
135 "Number of aio requests presently handled by the buf subsystem");
137 /* Number of async I/O thread in the process of being started */
138 /* XXX This should be local to _aio_aqueue() */
139 static int num_aio_resv_start
= 0;
141 static int aiod_timeout
;
142 SYSCTL_INT(_vfs_aio
, OID_AUTO
, aiod_timeout
, CTLFLAG_RW
, &aiod_timeout
, 0,
143 "Timeout value for synchronous aio operations");
145 static int aiod_lifetime
;
146 SYSCTL_INT(_vfs_aio
, OID_AUTO
, aiod_lifetime
, CTLFLAG_RW
, &aiod_lifetime
, 0,
147 "Maximum lifetime for idle aiod");
149 static int max_aio_per_proc
= MAX_AIO_PER_PROC
;
150 SYSCTL_INT(_vfs_aio
, OID_AUTO
, max_aio_per_proc
, CTLFLAG_RW
, &max_aio_per_proc
,
151 0, "Maximum active aio requests per process (stored in the process)");
153 static int max_aio_queue_per_proc
= MAX_AIO_QUEUE_PER_PROC
;
154 SYSCTL_INT(_vfs_aio
, OID_AUTO
, max_aio_queue_per_proc
, CTLFLAG_RW
,
155 &max_aio_queue_per_proc
, 0,
156 "Maximum queued aio requests per process (stored in the process)");
158 static int max_buf_aio
= MAX_BUF_AIO
;
159 SYSCTL_INT(_vfs_aio
, OID_AUTO
, max_buf_aio
, CTLFLAG_RW
, &max_buf_aio
, 0,
160 "Maximum buf aio requests per process (stored in the process)");
165 #define AIOP_FREE 0x1 /* proc on free queue */
166 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
169 int aioprocflags
; /* AIO proc flags */
170 TAILQ_ENTRY(aioproclist
) list
; /* List of processes */
171 struct proc
*aioproc
; /* The AIO thread */
175 * data-structure for lio signal management
179 int lioj_buffer_count
;
180 int lioj_buffer_finished_count
;
181 int lioj_queue_count
;
182 int lioj_queue_finished_count
;
183 struct sigevent lioj_signal
; /* signal on all I/O done */
184 TAILQ_ENTRY(aio_liojob
) lioj_list
;
185 struct kaioinfo
*lioj_ki
;
187 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
188 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
191 * per process aio data structure
194 int kaio_flags
; /* per process kaio flags */
195 int kaio_maxactive_count
; /* maximum number of AIOs */
196 int kaio_active_count
; /* number of currently used AIOs */
197 int kaio_qallowed_count
; /* maxiumu size of AIO queue */
198 int kaio_queue_count
; /* size of AIO queue */
199 int kaio_ballowed_count
; /* maximum number of buffers */
200 int kaio_queue_finished_count
; /* number of daemon jobs finished */
201 int kaio_buffer_count
; /* number of physio buffers */
202 int kaio_buffer_finished_count
; /* count of I/O done */
203 struct proc
*kaio_p
; /* process that uses this kaio block */
204 TAILQ_HEAD(,aio_liojob
) kaio_liojoblist
; /* list of lio jobs */
205 TAILQ_HEAD(,aiocblist
) kaio_jobqueue
; /* job queue for process */
206 TAILQ_HEAD(,aiocblist
) kaio_jobdone
; /* done queue for process */
207 TAILQ_HEAD(,aiocblist
) kaio_bufqueue
; /* buffer job queue for process */
208 TAILQ_HEAD(,aiocblist
) kaio_bufdone
; /* buffer done queue for process */
209 TAILQ_HEAD(,aiocblist
) kaio_sockqueue
; /* queue for aios waiting on sockets */
212 #define KAIO_RUNDOWN 0x1 /* process is being run down */
213 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
215 static TAILQ_HEAD(,aioproclist
) aio_freeproc
, aio_activeproc
;
216 static TAILQ_HEAD(,aiocblist
) aio_jobs
; /* Async job list */
217 static TAILQ_HEAD(,aiocblist
) aio_bufjobs
; /* Phys I/O job list */
218 static TAILQ_HEAD(,aiocblist
) aio_freejobs
; /* Pool of free jobs */
220 static void aio_init_aioinfo(struct proc
*p
);
221 static void aio_onceonly(void *);
222 static int aio_free_entry(struct aiocblist
*aiocbe
);
223 static void aio_process(struct aiocblist
*aiocbe
);
224 static int aio_newproc(void);
225 static int aio_aqueue(struct aiocb
*job
, int type
);
226 static void aio_physwakeup(struct bio
*bio
);
227 static int aio_fphysio(struct aiocblist
*aiocbe
);
228 static int aio_qphysio(struct proc
*p
, struct aiocblist
*iocb
);
229 static void aio_daemon(void *uproc
, struct trapframe
*frame
);
230 static void process_signal(void *aioj
);
232 SYSINIT(aio
, SI_SUB_VFS
, SI_ORDER_ANY
, aio_onceonly
, NULL
);
236 * kaio Per process async io info
237 * aiop async io thread data
238 * aiocb async io jobs
239 * aiol list io job pointer - internal to aio_suspend XXX
240 * aiolio list io jobs
242 static vm_zone_t kaio_zone
, aiop_zone
, aiocb_zone
, aiol_zone
, aiolio_zone
;
245 * Startup initialization
248 aio_onceonly(void *na
)
250 TAILQ_INIT(&aio_freeproc
);
251 TAILQ_INIT(&aio_activeproc
);
252 TAILQ_INIT(&aio_jobs
);
253 TAILQ_INIT(&aio_bufjobs
);
254 TAILQ_INIT(&aio_freejobs
);
255 kaio_zone
= zinit("AIO", sizeof(struct kaioinfo
), 0, 0, 1);
256 aiop_zone
= zinit("AIOP", sizeof(struct aioproclist
), 0, 0, 1);
257 aiocb_zone
= zinit("AIOCB", sizeof(struct aiocblist
), 0, 0, 1);
258 aiol_zone
= zinit("AIOL", AIO_LISTIO_MAX
*sizeof(intptr_t), 0, 0, 1);
259 aiolio_zone
= zinit("AIOLIO", sizeof(struct aio_liojob
), 0, 0, 1);
260 aiod_timeout
= AIOD_TIMEOUT_DEFAULT
;
261 aiod_lifetime
= AIOD_LIFETIME_DEFAULT
;
266 * Init the per-process aioinfo structure. The aioinfo limits are set
267 * per-process for user limit (resource) management.
270 aio_init_aioinfo(struct proc
*p
)
273 if (p
->p_aioinfo
== NULL
) {
274 ki
= zalloc(kaio_zone
);
277 ki
->kaio_maxactive_count
= max_aio_per_proc
;
278 ki
->kaio_active_count
= 0;
279 ki
->kaio_qallowed_count
= max_aio_queue_per_proc
;
280 ki
->kaio_queue_count
= 0;
281 ki
->kaio_ballowed_count
= max_buf_aio
;
282 ki
->kaio_buffer_count
= 0;
283 ki
->kaio_buffer_finished_count
= 0;
285 TAILQ_INIT(&ki
->kaio_jobdone
);
286 TAILQ_INIT(&ki
->kaio_jobqueue
);
287 TAILQ_INIT(&ki
->kaio_bufdone
);
288 TAILQ_INIT(&ki
->kaio_bufqueue
);
289 TAILQ_INIT(&ki
->kaio_liojoblist
);
290 TAILQ_INIT(&ki
->kaio_sockqueue
);
293 while (num_aio_procs
< target_aio_procs
)
298 * Free a job entry. Wait for completion if it is currently active, but don't
299 * delay forever. If we delay, we return a flag that says that we have to
300 * restart the queue scan.
303 aio_free_entry(struct aiocblist
*aiocbe
)
306 struct aio_liojob
*lj
;
310 if (aiocbe
->jobstate
== JOBST_NULL
)
311 panic("aio_free_entry: freeing already free job");
313 p
= aiocbe
->userproc
;
317 panic("aio_free_entry: missing p->p_aioinfo");
319 while (aiocbe
->jobstate
== JOBST_JOBRUNNING
) {
320 aiocbe
->jobflags
|= AIOCBLIST_RUNDOWN
;
321 tsleep(aiocbe
, 0, "jobwai", 0);
323 if (aiocbe
->bp
== NULL
) {
324 if (ki
->kaio_queue_count
<= 0)
325 panic("aio_free_entry: process queue size <= 0");
326 if (num_queue_count
<= 0)
327 panic("aio_free_entry: system wide queue size <= 0");
330 lj
->lioj_queue_count
--;
331 if (aiocbe
->jobflags
& AIOCBLIST_DONE
)
332 lj
->lioj_queue_finished_count
--;
334 ki
->kaio_queue_count
--;
335 if (aiocbe
->jobflags
& AIOCBLIST_DONE
)
336 ki
->kaio_queue_finished_count
--;
340 lj
->lioj_buffer_count
--;
341 if (aiocbe
->jobflags
& AIOCBLIST_DONE
)
342 lj
->lioj_buffer_finished_count
--;
344 if (aiocbe
->jobflags
& AIOCBLIST_DONE
)
345 ki
->kaio_buffer_finished_count
--;
346 ki
->kaio_buffer_count
--;
350 /* aiocbe is going away, we need to destroy any knotes */
351 /* XXX lwp knote wants a thread, but only cares about the process */
352 knote_remove(FIRST_LWP_IN_PROC(p
)->lwp_thread
, &aiocbe
->klist
);
354 if ((ki
->kaio_flags
& KAIO_WAKEUP
) || ((ki
->kaio_flags
& KAIO_RUNDOWN
)
355 && ((ki
->kaio_buffer_count
== 0) && (ki
->kaio_queue_count
== 0)))) {
356 ki
->kaio_flags
&= ~KAIO_WAKEUP
;
360 if (aiocbe
->jobstate
== JOBST_JOBQBUF
) {
361 if ((error
= aio_fphysio(aiocbe
)) != 0)
363 if (aiocbe
->jobstate
!= JOBST_JOBBFINISHED
)
364 panic("aio_free_entry: invalid physio finish-up state");
366 TAILQ_REMOVE(&ki
->kaio_bufdone
, aiocbe
, plist
);
368 } else if (aiocbe
->jobstate
== JOBST_JOBQGLOBAL
) {
370 TAILQ_REMOVE(&aio_jobs
, aiocbe
, list
);
371 TAILQ_REMOVE(&ki
->kaio_jobqueue
, aiocbe
, plist
);
373 } else if (aiocbe
->jobstate
== JOBST_JOBFINISHED
)
374 TAILQ_REMOVE(&ki
->kaio_jobdone
, aiocbe
, plist
);
375 else if (aiocbe
->jobstate
== JOBST_JOBBFINISHED
) {
377 TAILQ_REMOVE(&ki
->kaio_bufdone
, aiocbe
, plist
);
380 vunmapbuf(aiocbe
->bp
);
381 relpbuf(aiocbe
->bp
, NULL
);
385 if (lj
&& (lj
->lioj_buffer_count
== 0) && (lj
->lioj_queue_count
== 0)) {
386 TAILQ_REMOVE(&ki
->kaio_liojoblist
, lj
, lioj_list
);
387 zfree(aiolio_zone
, lj
);
389 aiocbe
->jobstate
= JOBST_NULL
;
390 callout_stop(&aiocbe
->timeout
);
391 fdrop(aiocbe
->fd_file
);
392 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
398 * Rundown the jobs for a given process.
401 aio_proc_rundown(struct proc
*p
)
407 struct aio_liojob
*lj
, *ljn
;
408 struct aiocblist
*aiocbe
, *aiocbn
;
416 ki
->kaio_flags
|= LIOJ_SIGNAL_POSTED
;
417 while ((ki
->kaio_active_count
> 0) || (ki
->kaio_buffer_count
>
418 ki
->kaio_buffer_finished_count
)) {
419 ki
->kaio_flags
|= KAIO_RUNDOWN
;
420 if (tsleep(p
, 0, "kaiowt", aiod_timeout
))
425 * Move any aio ops that are waiting on socket I/O to the normal job
426 * queues so they are cleaned up with any others.
429 for (aiocbe
= TAILQ_FIRST(&ki
->kaio_sockqueue
); aiocbe
; aiocbe
=
431 aiocbn
= TAILQ_NEXT(aiocbe
, plist
);
432 fp
= aiocbe
->fd_file
;
434 so
= (struct socket
*)fp
->f_data
;
435 TAILQ_REMOVE(&so
->so_aiojobq
, aiocbe
, list
);
436 if (TAILQ_EMPTY(&so
->so_aiojobq
)) {
437 so
->so_snd
.ssb_flags
&= ~SSB_AIO
;
438 so
->so_rcv
.ssb_flags
&= ~SSB_AIO
;
441 TAILQ_REMOVE(&ki
->kaio_sockqueue
, aiocbe
, plist
);
442 TAILQ_INSERT_HEAD(&aio_jobs
, aiocbe
, list
);
443 TAILQ_INSERT_HEAD(&ki
->kaio_jobqueue
, aiocbe
, plist
);
448 for (aiocbe
= TAILQ_FIRST(&ki
->kaio_jobdone
); aiocbe
; aiocbe
= aiocbn
) {
449 aiocbn
= TAILQ_NEXT(aiocbe
, plist
);
450 if (aio_free_entry(aiocbe
))
455 for (aiocbe
= TAILQ_FIRST(&ki
->kaio_jobqueue
); aiocbe
; aiocbe
=
457 aiocbn
= TAILQ_NEXT(aiocbe
, plist
);
458 if (aio_free_entry(aiocbe
))
464 while (TAILQ_FIRST(&ki
->kaio_bufqueue
)) {
465 ki
->kaio_flags
|= KAIO_WAKEUP
;
466 tsleep(p
, 0, "aioprn", 0);
474 for (aiocbe
= TAILQ_FIRST(&ki
->kaio_bufdone
); aiocbe
; aiocbe
= aiocbn
) {
475 aiocbn
= TAILQ_NEXT(aiocbe
, plist
);
476 if (aio_free_entry(aiocbe
)) {
484 * If we've slept, jobs might have moved from one queue to another.
485 * Retry rundown if we didn't manage to empty the queues.
487 if (TAILQ_FIRST(&ki
->kaio_jobdone
) != NULL
||
488 TAILQ_FIRST(&ki
->kaio_jobqueue
) != NULL
||
489 TAILQ_FIRST(&ki
->kaio_bufqueue
) != NULL
||
490 TAILQ_FIRST(&ki
->kaio_bufdone
) != NULL
)
493 for (lj
= TAILQ_FIRST(&ki
->kaio_liojoblist
); lj
; lj
= ljn
) {
494 ljn
= TAILQ_NEXT(lj
, lioj_list
);
495 if ((lj
->lioj_buffer_count
== 0) && (lj
->lioj_queue_count
==
497 TAILQ_REMOVE(&ki
->kaio_liojoblist
, lj
, lioj_list
);
498 zfree(aiolio_zone
, lj
);
501 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
502 "QF:%d\n", lj
->lioj_buffer_count
,
503 lj
->lioj_buffer_finished_count
,
504 lj
->lioj_queue_count
,
505 lj
->lioj_queue_finished_count
);
510 zfree(kaio_zone
, ki
);
517 * Select a job to run (called by an AIO daemon).
519 static struct aiocblist
*
520 aio_selectjob(struct aioproclist
*aiop
)
522 struct aiocblist
*aiocbe
;
527 for (aiocbe
= TAILQ_FIRST(&aio_jobs
); aiocbe
; aiocbe
=
528 TAILQ_NEXT(aiocbe
, list
)) {
529 userp
= aiocbe
->userproc
;
530 ki
= userp
->p_aioinfo
;
532 if (ki
->kaio_active_count
< ki
->kaio_maxactive_count
) {
533 TAILQ_REMOVE(&aio_jobs
, aiocbe
, list
);
544 * The AIO processing activity. This is the code that does the I/O request for
545 * the non-physio version of the operations. The normal vn operations are used,
546 * and this code should work in all instances for every type of file, including
547 * pipes, sockets, fifos, and regular files.
550 aio_process(struct aiocblist
*aiocbe
)
559 int oublock_st
, oublock_end
;
560 int inblock_st
, inblock_end
;
563 cb
= &aiocbe
->uaiocb
;
564 fp
= aiocbe
->fd_file
;
566 aiov
.iov_base
= (void *)(uintptr_t)cb
->aio_buf
;
567 aiov
.iov_len
= cb
->aio_nbytes
;
569 auio
.uio_iov
= &aiov
;
571 auio
.uio_offset
= cb
->aio_offset
;
572 auio
.uio_resid
= cb
->aio_nbytes
;
573 cnt
= cb
->aio_nbytes
;
574 auio
.uio_segflg
= UIO_USERSPACE
;
577 inblock_st
= mytd
->td_lwp
->lwp_ru
.ru_inblock
;
578 oublock_st
= mytd
->td_lwp
->lwp_ru
.ru_oublock
;
580 * _aio_aqueue() acquires a reference to the file that is
581 * released in aio_free_entry().
583 if (cb
->aio_lio_opcode
== LIO_READ
) {
584 auio
.uio_rw
= UIO_READ
;
585 error
= fo_read(fp
, &auio
, fp
->f_cred
, O_FOFFSET
);
587 auio
.uio_rw
= UIO_WRITE
;
588 error
= fo_write(fp
, &auio
, fp
->f_cred
, O_FOFFSET
);
590 inblock_end
= mytd
->td_lwp
->lwp_ru
.ru_inblock
;
591 oublock_end
= mytd
->td_lwp
->lwp_ru
.ru_oublock
;
593 aiocbe
->inputcharge
= inblock_end
- inblock_st
;
594 aiocbe
->outputcharge
= oublock_end
- oublock_st
;
596 if ((error
) && (auio
.uio_resid
!= cnt
)) {
597 if (error
== ERESTART
|| error
== EINTR
|| error
== EWOULDBLOCK
)
599 if ((error
== EPIPE
) && (cb
->aio_lio_opcode
== LIO_WRITE
))
600 ksignal(aiocbe
->userproc
, SIGPIPE
);
603 cnt
-= auio
.uio_resid
;
604 cb
->_aiocb_private
.error
= error
;
605 cb
->_aiocb_private
.status
= cnt
;
609 * The AIO daemon, most of the actual work is done in aio_process,
610 * but the setup (and address space mgmt) is done in this routine.
612 * The MP lock is held on entry.
615 aio_daemon(void *uproc
, struct trapframe
*frame
)
617 struct aio_liojob
*lj
;
619 struct aiocblist
*aiocbe
;
620 struct aioproclist
*aiop
;
622 struct proc
*mycp
, *userp
;
623 struct vmspace
*curvm
;
627 mylwp
= curthread
->td_lwp
;
628 mycp
= mylwp
->lwp_proc
;
630 if (mycp
->p_textvp
) {
631 vrele(mycp
->p_textvp
);
632 mycp
->p_textvp
= NULL
;
636 * Allocate and ready the aio control info. There is one aiop structure
639 aiop
= zalloc(aiop_zone
);
640 aiop
->aioproc
= mycp
;
641 aiop
->aioprocflags
|= AIOP_FREE
;
646 * Place thread (lightweight process) onto the AIO free thread list.
648 if (TAILQ_EMPTY(&aio_freeproc
))
649 wakeup(&aio_freeproc
);
650 TAILQ_INSERT_HEAD(&aio_freeproc
, aiop
, list
);
654 /* Make up a name for the daemon. */
655 strcpy(mycp
->p_comm
, "aiod");
658 * Get rid of our current filedescriptors. AIOD's don't need any
659 * filedescriptors, except as temporarily inherited from the client.
660 * Credentials are also cloned, and made equivalent to "root".
663 cr
= cratom(&mycp
->p_ucred
);
665 uireplace(&cr
->cr_uidinfo
, uifind(0));
667 cr
->cr_groups
[0] = 1;
669 /* The daemon resides in its own pgrp. */
670 enterpgrp(mycp
, mycp
->p_pid
, 1);
672 /* Mark special process type. */
673 mycp
->p_flag
|= P_SYSTEM
| P_KTHREADP
;
676 * Wakeup parent process. (Parent sleeps to keep from blasting away
677 * and creating too many daemons.)
684 * Take daemon off of free queue
686 if (aiop
->aioprocflags
& AIOP_FREE
) {
688 TAILQ_REMOVE(&aio_freeproc
, aiop
, list
);
689 TAILQ_INSERT_TAIL(&aio_activeproc
, aiop
, list
);
690 aiop
->aioprocflags
&= ~AIOP_FREE
;
693 aiop
->aioprocflags
&= ~AIOP_SCHED
;
698 while ((aiocbe
= aio_selectjob(aiop
)) != NULL
) {
699 cb
= &aiocbe
->uaiocb
;
700 userp
= aiocbe
->userproc
;
702 aiocbe
->jobstate
= JOBST_JOBRUNNING
;
705 * Connect to process address space for user program.
707 if (curvm
!= userp
->p_vmspace
) {
708 pmap_setlwpvm(mylwp
, userp
->p_vmspace
);
710 sysref_put(&curvm
->vm_sysref
);
711 curvm
= userp
->p_vmspace
;
712 sysref_get(&curvm
->vm_sysref
);
715 ki
= userp
->p_aioinfo
;
718 /* Account for currently active jobs. */
719 ki
->kaio_active_count
++;
721 /* Do the I/O function. */
724 /* Decrement the active job count. */
725 ki
->kaio_active_count
--;
728 * Increment the completion count for wakeup/signal
731 aiocbe
->jobflags
|= AIOCBLIST_DONE
;
732 ki
->kaio_queue_finished_count
++;
734 lj
->lioj_queue_finished_count
++;
735 if ((ki
->kaio_flags
& KAIO_WAKEUP
) || ((ki
->kaio_flags
736 & KAIO_RUNDOWN
) && (ki
->kaio_active_count
== 0))) {
737 ki
->kaio_flags
&= ~KAIO_WAKEUP
;
742 if (lj
&& (lj
->lioj_flags
&
743 (LIOJ_SIGNAL
|LIOJ_SIGNAL_POSTED
)) == LIOJ_SIGNAL
) {
744 if ((lj
->lioj_queue_finished_count
==
745 lj
->lioj_queue_count
) &&
746 (lj
->lioj_buffer_finished_count
==
747 lj
->lioj_buffer_count
)) {
749 lj
->lioj_signal
.sigev_signo
);
756 aiocbe
->jobstate
= JOBST_JOBFINISHED
;
759 TAILQ_REMOVE(&ki
->kaio_jobqueue
, aiocbe
, plist
);
760 TAILQ_INSERT_TAIL(&ki
->kaio_jobdone
, aiocbe
, plist
);
762 KNOTE(&aiocbe
->klist
, 0);
764 if (aiocbe
->jobflags
& AIOCBLIST_RUNDOWN
) {
766 aiocbe
->jobflags
&= ~AIOCBLIST_RUNDOWN
;
769 if (cb
->aio_sigevent
.sigev_notify
== SIGEV_SIGNAL
) {
770 ksignal(userp
, cb
->aio_sigevent
.sigev_signo
);
775 * Disconnect from user address space.
778 /* swap our original address space back in */
779 pmap_setlwpvm(mylwp
, mycp
->p_vmspace
);
780 sysref_put(&curvm
->vm_sysref
);
785 * If we are the first to be put onto the free queue, wakeup
786 * anyone waiting for a daemon.
789 TAILQ_REMOVE(&aio_activeproc
, aiop
, list
);
790 if (TAILQ_EMPTY(&aio_freeproc
))
791 wakeup(&aio_freeproc
);
792 TAILQ_INSERT_HEAD(&aio_freeproc
, aiop
, list
);
793 aiop
->aioprocflags
|= AIOP_FREE
;
797 * If daemon is inactive for a long time, allow it to exit,
798 * thereby freeing resources.
800 if (((aiop
->aioprocflags
& AIOP_SCHED
) == 0) && tsleep(mycp
,
801 0, "aiordy", aiod_lifetime
)) {
803 if (TAILQ_EMPTY(&aio_jobs
)) {
804 if ((aiop
->aioprocflags
& AIOP_FREE
) &&
805 (num_aio_procs
> target_aio_procs
)) {
806 TAILQ_REMOVE(&aio_freeproc
, aiop
, list
);
808 zfree(aiop_zone
, aiop
);
811 if (mycp
->p_vmspace
->vm_sysref
.refcnt
<= 1) {
812 kprintf("AIOD: bad vm refcnt for"
813 " exiting daemon: %d\n",
814 mycp
->p_vmspace
->vm_sysref
.refcnt
);
826 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
827 * AIO daemon modifies its environment itself.
833 struct lwp
*lp
, *nlp
;
837 error
= fork1(lp
, RFPROC
|RFMEM
|RFNOWAIT
, &np
);
840 nlp
= ONLY_LWP_IN_PROC(np
);
841 cpu_set_fork_handler(nlp
, aio_daemon
, curproc
);
842 start_forked_proc(lp
, np
);
845 * Wait until daemon is started, but continue on just in case to
846 * handle error conditions.
848 error
= tsleep(np
, 0, "aiosta", aiod_timeout
);
855 * Try the high-performance, low-overhead physio method for eligible
856 * VCHR devices. This method doesn't use an aio helper thread, and
857 * thus has very low overhead.
859 * Assumes that the caller, _aio_aqueue(), has incremented the file
860 * structure's reference count, preventing its deallocation for the
861 * duration of this call.
864 aio_qphysio(struct proc
*p
, struct aiocblist
*aiocbe
)
872 struct aio_liojob
*lj
;
875 cb
= &aiocbe
->uaiocb
;
876 fp
= aiocbe
->fd_file
;
878 if (fp
->f_type
!= DTYPE_VNODE
)
881 vp
= (struct vnode
*)fp
->f_data
;
884 * If its not a disk, we don't want to return a positive error.
885 * It causes the aio code to not fall through to try the thread
886 * way when you're talking to a regular file.
888 if (!vn_isdisk(vp
, &error
)) {
889 if (error
== ENOTBLK
)
895 if (cb
->aio_nbytes
% vp
->v_rdev
->si_bsize_phys
)
899 MAXPHYS
- (((vm_offset_t
) cb
->aio_buf
) & PAGE_MASK
))
903 if (ki
->kaio_buffer_count
>= ki
->kaio_ballowed_count
)
906 ki
->kaio_buffer_count
++;
910 lj
->lioj_buffer_count
++;
912 /* Create and build a buffer header for a transfer. */
917 * Get a copy of the kva from the physical buffer.
919 bp
->b_bio1
.bio_caller_info1
.ptr
= p
;
922 bp
->b_cmd
= (cb
->aio_lio_opcode
== LIO_WRITE
) ?
923 BUF_CMD_WRITE
: BUF_CMD_READ
;
924 bp
->b_bio1
.bio_done
= aio_physwakeup
;
925 bp
->b_bio1
.bio_flags
|= BIO_SYNC
;
926 bp
->b_bio1
.bio_offset
= cb
->aio_offset
;
928 /* Bring buffer into kernel space. */
929 if (vmapbuf(bp
, __DEVOLATILE(char *, cb
->aio_buf
), cb
->aio_nbytes
) < 0) {
937 bp
->b_bio1
.bio_caller_info2
.ptr
= aiocbe
;
938 TAILQ_INSERT_TAIL(&aio_bufjobs
, aiocbe
, list
);
939 TAILQ_INSERT_TAIL(&ki
->kaio_bufqueue
, aiocbe
, plist
);
940 aiocbe
->jobstate
= JOBST_JOBQBUF
;
941 cb
->_aiocb_private
.status
= cb
->aio_nbytes
;
948 * Perform the transfer. vn_strategy must be used even though we
949 * know we have a device in order to deal with requests which exceed
950 * device DMA limitations.
952 vn_strategy(vp
, &bp
->b_bio1
);
959 * If we had an error invoking the request, or an error in processing
960 * the request before we have returned, we process it as an error in
961 * transfer. Note that such an I/O error is not indicated immediately,
962 * but is returned using the aio_error mechanism. In this case,
963 * aio_suspend will return immediately.
965 if (bp
->b_error
|| (bp
->b_flags
& B_ERROR
)) {
966 struct aiocb
*job
= aiocbe
->uuaiocb
;
968 aiocbe
->uaiocb
._aiocb_private
.status
= 0;
969 suword(&job
->_aiocb_private
.status
, 0);
970 aiocbe
->uaiocb
._aiocb_private
.error
= bp
->b_error
;
971 suword(&job
->_aiocb_private
.error
, bp
->b_error
);
973 ki
->kaio_buffer_finished_count
++;
975 if (aiocbe
->jobstate
!= JOBST_JOBBFINISHED
) {
976 aiocbe
->jobstate
= JOBST_JOBBFINISHED
;
977 aiocbe
->jobflags
|= AIOCBLIST_DONE
;
978 TAILQ_REMOVE(&aio_bufjobs
, aiocbe
, list
);
979 TAILQ_REMOVE(&ki
->kaio_bufqueue
, aiocbe
, plist
);
980 TAILQ_INSERT_TAIL(&ki
->kaio_bufdone
, aiocbe
, plist
);
987 KNOTE(&aiocbe
->klist
, 0);
991 ki
->kaio_buffer_count
--;
993 lj
->lioj_buffer_count
--;
1000 * This waits/tests physio completion.
1003 aio_fphysio(struct aiocblist
*iocb
)
1010 error
= biowait_timeout(&bp
->b_bio1
, "physstr", aiod_timeout
);
1011 if (error
== EWOULDBLOCK
)
1014 /* Release mapping into kernel space. */
1020 /* Check for an error. */
1021 if (bp
->b_flags
& B_ERROR
)
1022 error
= bp
->b_error
;
1027 #endif /* VFS_AIO */
1030 * Wake up aio requests that may be serviceable now.
1033 aio_swake(struct socket
*so
, struct signalsockbuf
*ssb
)
1038 struct aiocblist
*cb
,*cbn
;
1040 struct kaioinfo
*ki
= NULL
;
1041 int opcode
, wakecount
= 0;
1042 struct aioproclist
*aiop
;
1044 if (ssb
== &so
->so_snd
) {
1046 so
->so_snd
.ssb_flags
&= ~SSB_AIO
;
1049 so
->so_rcv
.ssb_flags
&= ~SSB_AIO
;
1052 for (cb
= TAILQ_FIRST(&so
->so_aiojobq
); cb
; cb
= cbn
) {
1053 cbn
= TAILQ_NEXT(cb
, list
);
1054 if (opcode
== cb
->uaiocb
.aio_lio_opcode
) {
1057 TAILQ_REMOVE(&so
->so_aiojobq
, cb
, list
);
1058 TAILQ_REMOVE(&ki
->kaio_sockqueue
, cb
, plist
);
1059 TAILQ_INSERT_TAIL(&aio_jobs
, cb
, list
);
1060 TAILQ_INSERT_TAIL(&ki
->kaio_jobqueue
, cb
, plist
);
1062 if (cb
->jobstate
!= JOBST_JOBQGLOBAL
)
1063 panic("invalid queue value");
1067 while (wakecount
--) {
1068 if ((aiop
= TAILQ_FIRST(&aio_freeproc
)) != 0) {
1069 TAILQ_REMOVE(&aio_freeproc
, aiop
, list
);
1070 TAILQ_INSERT_TAIL(&aio_activeproc
, aiop
, list
);
1071 aiop
->aioprocflags
&= ~AIOP_FREE
;
1072 wakeup(aiop
->aioproc
);
1075 #endif /* VFS_AIO */
1080 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1081 * technique is done in this code.
1084 _aio_aqueue(struct aiocb
*job
, struct aio_liojob
*lj
, int type
)
1086 struct proc
*p
= curproc
;
1087 struct filedesc
*fdp
;
1092 int opcode
, user_opcode
;
1093 struct aiocblist
*aiocbe
;
1094 struct aioproclist
*aiop
;
1095 struct kaioinfo
*ki
;
1100 if ((aiocbe
= TAILQ_FIRST(&aio_freejobs
)) != NULL
)
1101 TAILQ_REMOVE(&aio_freejobs
, aiocbe
, list
);
1103 aiocbe
= zalloc (aiocb_zone
);
1105 aiocbe
->inputcharge
= 0;
1106 aiocbe
->outputcharge
= 0;
1107 callout_init(&aiocbe
->timeout
);
1108 SLIST_INIT(&aiocbe
->klist
);
1110 suword(&job
->_aiocb_private
.status
, -1);
1111 suword(&job
->_aiocb_private
.error
, 0);
1112 suword(&job
->_aiocb_private
.kernelinfo
, -1);
1114 error
= copyin(job
, &aiocbe
->uaiocb
, sizeof(aiocbe
->uaiocb
));
1116 suword(&job
->_aiocb_private
.error
, error
);
1117 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1120 if (aiocbe
->uaiocb
.aio_sigevent
.sigev_notify
== SIGEV_SIGNAL
&&
1121 !_SIG_VALID(aiocbe
->uaiocb
.aio_sigevent
.sigev_signo
)) {
1122 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1126 /* Save userspace address of the job info. */
1127 aiocbe
->uuaiocb
= job
;
1129 /* Get the opcode. */
1130 user_opcode
= aiocbe
->uaiocb
.aio_lio_opcode
;
1131 if (type
!= LIO_NOP
)
1132 aiocbe
->uaiocb
.aio_lio_opcode
= type
;
1133 opcode
= aiocbe
->uaiocb
.aio_lio_opcode
;
1135 /* Get the fd info for process. */
1139 * Range check file descriptor.
1141 fd
= aiocbe
->uaiocb
.aio_fildes
;
1142 if (fd
>= fdp
->fd_nfiles
) {
1143 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1145 suword(&job
->_aiocb_private
.error
, EBADF
);
1149 fp
= aiocbe
->fd_file
= fdp
->fd_files
[fd
].fp
;
1150 if ((fp
== NULL
) || ((opcode
== LIO_WRITE
) && ((fp
->f_flag
& FWRITE
) ==
1152 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1154 suword(&job
->_aiocb_private
.error
, EBADF
);
1159 if (aiocbe
->uaiocb
.aio_offset
== -1LL) {
1163 error
= suword(&job
->_aiocb_private
.kernelinfo
, jobrefid
);
1168 aiocbe
->uaiocb
._aiocb_private
.kernelinfo
= (void *)(intptr_t)jobrefid
;
1169 if (jobrefid
== LONG_MAX
)
1174 if (opcode
== LIO_NOP
) {
1176 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1178 suword(&job
->_aiocb_private
.error
, 0);
1179 suword(&job
->_aiocb_private
.status
, 0);
1180 suword(&job
->_aiocb_private
.kernelinfo
, 0);
1184 if ((opcode
!= LIO_READ
) && (opcode
!= LIO_WRITE
)) {
1186 suword(&job
->_aiocb_private
.status
, 0);
1191 if (aiocbe
->uaiocb
.aio_sigevent
.sigev_notify
== SIGEV_KEVENT
) {
1192 kev
.ident
= aiocbe
->uaiocb
.aio_sigevent
.sigev_notify_kqueue
;
1193 kev
.udata
= aiocbe
->uaiocb
.aio_sigevent
.sigev_value
.sigval_ptr
;
1197 * This method for requesting kevent-based notification won't
1198 * work on the alpha, since we're passing in a pointer
1199 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1200 * based method instead.
1202 if (user_opcode
== LIO_NOP
|| user_opcode
== LIO_READ
||
1203 user_opcode
== LIO_WRITE
)
1206 error
= copyin((struct kevent
*)(uintptr_t)user_opcode
,
1211 if ((u_int
)kev
.ident
>= fdp
->fd_nfiles
||
1212 (kq_fp
= fdp
->fd_files
[kev
.ident
].fp
) == NULL
||
1213 (kq_fp
->f_type
!= DTYPE_KQUEUE
)) {
1217 kq
= (struct kqueue
*)kq_fp
->f_data
;
1218 kev
.ident
= (uintptr_t)aiocbe
->uuaiocb
;
1219 kev
.filter
= EVFILT_AIO
;
1220 kev
.flags
= EV_ADD
| EV_ENABLE
| EV_FLAG1
;
1221 kev
.data
= (intptr_t)aiocbe
;
1222 /* XXX lwp kqueue_register takes a thread, but only uses its proc */
1223 error
= kqueue_register(kq
, &kev
, FIRST_LWP_IN_PROC(p
)->lwp_thread
);
1227 TAILQ_INSERT_HEAD(&aio_freejobs
, aiocbe
, list
);
1229 suword(&job
->_aiocb_private
.error
, error
);
1234 suword(&job
->_aiocb_private
.error
, EINPROGRESS
);
1235 aiocbe
->uaiocb
._aiocb_private
.error
= EINPROGRESS
;
1236 aiocbe
->userproc
= p
;
1237 aiocbe
->jobflags
= 0;
1241 if (fp
->f_type
== DTYPE_SOCKET
) {
1243 * Alternate queueing for socket ops: Reach down into the
1244 * descriptor to get the socket data. Then check to see if the
1245 * socket is ready to be read or written (based on the requested
1248 * If it is not ready for io, then queue the aiocbe on the
1249 * socket, and set the flags so we get a call when ssb_notify()
1252 so
= (struct socket
*)fp
->f_data
;
1254 if (((opcode
== LIO_READ
) && (!soreadable(so
))) || ((opcode
==
1255 LIO_WRITE
) && (!sowriteable(so
)))) {
1256 TAILQ_INSERT_TAIL(&so
->so_aiojobq
, aiocbe
, list
);
1257 TAILQ_INSERT_TAIL(&ki
->kaio_sockqueue
, aiocbe
, plist
);
1258 if (opcode
== LIO_READ
)
1259 so
->so_rcv
.ssb_flags
|= SSB_AIO
;
1261 so
->so_snd
.ssb_flags
|= SSB_AIO
;
1262 aiocbe
->jobstate
= JOBST_JOBQGLOBAL
; /* XXX */
1263 ki
->kaio_queue_count
++;
1272 if ((error
= aio_qphysio(p
, aiocbe
)) == 0)
1275 suword(&job
->_aiocb_private
.status
, 0);
1276 aiocbe
->uaiocb
._aiocb_private
.error
= error
;
1277 suword(&job
->_aiocb_private
.error
, error
);
1281 /* No buffer for daemon I/O. */
1284 ki
->kaio_queue_count
++;
1286 lj
->lioj_queue_count
++;
1288 TAILQ_INSERT_TAIL(&ki
->kaio_jobqueue
, aiocbe
, plist
);
1289 TAILQ_INSERT_TAIL(&aio_jobs
, aiocbe
, list
);
1291 aiocbe
->jobstate
= JOBST_JOBQGLOBAL
;
1297 * If we don't have a free AIO process, and we are below our quota, then
1298 * start one. Otherwise, depend on the subsequent I/O completions to
1299 * pick-up this job. If we don't successfully create the new process
1300 * (thread) due to resource issues, we return an error for now (EAGAIN),
1301 * which is likely not the correct thing to do.
1305 if ((aiop
= TAILQ_FIRST(&aio_freeproc
)) != NULL
) {
1306 TAILQ_REMOVE(&aio_freeproc
, aiop
, list
);
1307 TAILQ_INSERT_TAIL(&aio_activeproc
, aiop
, list
);
1308 aiop
->aioprocflags
&= ~AIOP_FREE
;
1309 wakeup(aiop
->aioproc
);
1310 } else if (((num_aio_resv_start
+ num_aio_procs
) < max_aio_procs
) &&
1311 ((ki
->kaio_active_count
+ num_aio_resv_start
) <
1312 ki
->kaio_maxactive_count
)) {
1313 num_aio_resv_start
++;
1314 if ((error
= aio_newproc()) == 0) {
1315 num_aio_resv_start
--;
1318 num_aio_resv_start
--;
1326 * This routine queues an AIO request, checking for quotas.
1329 aio_aqueue(struct aiocb
*job
, int type
)
1331 struct proc
*p
= curproc
;
1332 struct kaioinfo
*ki
;
1334 if (p
->p_aioinfo
== NULL
)
1335 aio_init_aioinfo(p
);
1337 if (num_queue_count
>= max_queue_count
)
1341 if (ki
->kaio_queue_count
>= ki
->kaio_qallowed_count
)
1344 return _aio_aqueue(job
, NULL
, type
);
1346 #endif /* VFS_AIO */
1349 * Support the aio_return system call, as a side-effect, kernel resources are
1353 sys_aio_return(struct aio_return_args
*uap
)
1358 struct proc
*p
= curproc
;
1359 struct lwp
*lp
= curthread
->td_lwp
;
1361 struct aiocblist
*cb
, *ncb
;
1363 struct kaioinfo
*ki
;
1371 jobref
= fuword(&ujob
->_aiocb_private
.kernelinfo
);
1372 if (jobref
== -1 || jobref
== 0)
1375 TAILQ_FOREACH(cb
, &ki
->kaio_jobdone
, plist
) {
1376 if (((intptr_t) cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1378 if (ujob
== cb
->uuaiocb
) {
1379 uap
->sysmsg_result
=
1380 cb
->uaiocb
._aiocb_private
.status
;
1382 uap
->sysmsg_result
= EFAULT
;
1383 if (cb
->uaiocb
.aio_lio_opcode
== LIO_WRITE
) {
1384 lp
->lwp_ru
.ru_oublock
+= cb
->outputcharge
;
1385 cb
->outputcharge
= 0;
1386 } else if (cb
->uaiocb
.aio_lio_opcode
== LIO_READ
) {
1387 lp
->lwp_ru
.ru_inblock
+= cb
->inputcharge
;
1388 cb
->inputcharge
= 0;
1395 for (cb
= TAILQ_FIRST(&ki
->kaio_bufdone
); cb
; cb
= ncb
) {
1396 ncb
= TAILQ_NEXT(cb
, plist
);
1397 if (((intptr_t) cb
->uaiocb
._aiocb_private
.kernelinfo
)
1400 if (ujob
== cb
->uuaiocb
) {
1401 uap
->sysmsg_result
=
1402 cb
->uaiocb
._aiocb_private
.status
;
1404 uap
->sysmsg_result
= EFAULT
;
1412 #endif /* VFS_AIO */
1416 * Allow a process to wakeup when any of the I/O requests are completed.
1419 sys_aio_suspend(struct aio_suspend_args
*uap
)
1424 struct proc
*p
= curproc
;
1427 struct aiocb
*const *cbptr
, *cbp
;
1428 struct kaioinfo
*ki
;
1429 struct aiocblist
*cb
;
1434 struct aiocb
**ujoblist
;
1436 if (uap
->nent
> AIO_LISTIO_MAX
)
1441 /* Get timespec struct. */
1442 if ((error
= copyin(uap
->timeout
, &ts
, sizeof(ts
))) != 0)
1445 if (ts
.tv_nsec
< 0 || ts
.tv_nsec
>= 1000000000)
1448 TIMESPEC_TO_TIMEVAL(&atv
, &ts
);
1449 if (itimerfix(&atv
))
1451 timo
= tvtohz_high(&atv
);
1459 ijoblist
= zalloc(aiol_zone
);
1460 ujoblist
= zalloc(aiol_zone
);
1461 cbptr
= uap
->aiocbp
;
1463 for (i
= 0; i
< uap
->nent
; i
++) {
1464 cbp
= (struct aiocb
*)(intptr_t)fuword(&cbptr
[i
]);
1467 ujoblist
[njoblist
] = cbp
;
1468 ijoblist
[njoblist
] = fuword(&cbp
->_aiocb_private
.kernelinfo
);
1472 if (njoblist
== 0) {
1473 zfree(aiol_zone
, ijoblist
);
1474 zfree(aiol_zone
, ujoblist
);
1480 TAILQ_FOREACH(cb
, &ki
->kaio_jobdone
, plist
) {
1481 for (i
= 0; i
< njoblist
; i
++) {
1483 cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1485 if (ujoblist
[i
] != cb
->uuaiocb
)
1487 zfree(aiol_zone
, ijoblist
);
1488 zfree(aiol_zone
, ujoblist
);
1495 for (cb
= TAILQ_FIRST(&ki
->kaio_bufdone
); cb
; cb
=
1496 TAILQ_NEXT(cb
, plist
)) {
1497 for (i
= 0; i
< njoblist
; i
++) {
1499 cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1502 if (ujoblist
[i
] != cb
->uuaiocb
)
1504 zfree(aiol_zone
, ijoblist
);
1505 zfree(aiol_zone
, ujoblist
);
1511 ki
->kaio_flags
|= KAIO_WAKEUP
;
1512 error
= tsleep(p
, PCATCH
, "aiospn", timo
);
1515 if (error
== ERESTART
|| error
== EINTR
) {
1516 zfree(aiol_zone
, ijoblist
);
1517 zfree(aiol_zone
, ujoblist
);
1519 } else if (error
== EWOULDBLOCK
) {
1520 zfree(aiol_zone
, ijoblist
);
1521 zfree(aiol_zone
, ujoblist
);
1528 #endif /* VFS_AIO */
1532 * aio_cancel cancels any non-physio aio operations not currently in
1536 sys_aio_cancel(struct aio_cancel_args
*uap
)
1541 struct proc
*p
= curproc
;
1542 struct kaioinfo
*ki
;
1543 struct aiocblist
*cbe
, *cbn
;
1545 struct filedesc
*fdp
;
1554 if ((u_int
)uap
->fd
>= fdp
->fd_nfiles
||
1555 (fp
= fdp
->fd_files
[uap
->fd
].fp
) == NULL
)
1558 if (fp
->f_type
== DTYPE_VNODE
) {
1559 vp
= (struct vnode
*)fp
->f_data
;
1561 if (vn_isdisk(vp
,&error
)) {
1562 uap
->sysmsg_result
= AIO_NOTCANCELED
;
1565 } else if (fp
->f_type
== DTYPE_SOCKET
) {
1566 so
= (struct socket
*)fp
->f_data
;
1570 for (cbe
= TAILQ_FIRST(&so
->so_aiojobq
); cbe
; cbe
= cbn
) {
1571 cbn
= TAILQ_NEXT(cbe
, list
);
1572 if ((uap
->aiocbp
== NULL
) ||
1573 (uap
->aiocbp
== cbe
->uuaiocb
) ) {
1576 TAILQ_REMOVE(&so
->so_aiojobq
, cbe
, list
);
1577 TAILQ_REMOVE(&ki
->kaio_sockqueue
, cbe
, plist
);
1578 TAILQ_INSERT_TAIL(&ki
->kaio_jobdone
, cbe
, plist
);
1579 if (ki
->kaio_flags
& KAIO_WAKEUP
) {
1582 cbe
->jobstate
= JOBST_JOBFINISHED
;
1583 cbe
->uaiocb
._aiocb_private
.status
=-1;
1584 cbe
->uaiocb
._aiocb_private
.error
=ECANCELED
;
1586 /* XXX cancelled, knote? */
1587 if (cbe
->uaiocb
.aio_sigevent
.sigev_notify
==
1589 ksignal(cbe
->userproc
, cbe
->uaiocb
.aio_sigevent
.sigev_signo
);
1596 if ((cancelled
) && (uap
->aiocbp
)) {
1597 uap
->sysmsg_result
= AIO_CANCELED
;
1606 for (cbe
= TAILQ_FIRST(&ki
->kaio_jobqueue
); cbe
; cbe
= cbn
) {
1607 cbn
= TAILQ_NEXT(cbe
, plist
);
1609 if ((uap
->fd
== cbe
->uaiocb
.aio_fildes
) &&
1610 ((uap
->aiocbp
== NULL
) ||
1611 (uap
->aiocbp
== cbe
->uuaiocb
))) {
1613 if (cbe
->jobstate
== JOBST_JOBQGLOBAL
) {
1614 TAILQ_REMOVE(&aio_jobs
, cbe
, list
);
1615 TAILQ_REMOVE(&ki
->kaio_jobqueue
, cbe
, plist
);
1616 TAILQ_INSERT_TAIL(&ki
->kaio_jobdone
, cbe
,
1619 ki
->kaio_queue_finished_count
++;
1620 cbe
->jobstate
= JOBST_JOBFINISHED
;
1621 cbe
->uaiocb
._aiocb_private
.status
= -1;
1622 cbe
->uaiocb
._aiocb_private
.error
= ECANCELED
;
1623 /* XXX cancelled, knote? */
1624 if (cbe
->uaiocb
.aio_sigevent
.sigev_notify
==
1626 ksignal(cbe
->userproc
, cbe
->uaiocb
.aio_sigevent
.sigev_signo
);
1635 uap
->sysmsg_result
= AIO_NOTCANCELED
;
1639 uap
->sysmsg_result
= AIO_CANCELED
;
1642 uap
->sysmsg_result
= AIO_ALLDONE
;
1645 #endif /* VFS_AIO */
1649 * aio_error is implemented in the kernel level for compatibility purposes only.
1650 * For a user mode async implementation, it would be best to do it in a userland
1654 sys_aio_error(struct aio_error_args
*uap
)
1659 struct proc
*p
= curproc
;
1660 struct aiocblist
*cb
;
1661 struct kaioinfo
*ki
;
1668 jobref
= fuword(&uap
->aiocbp
->_aiocb_private
.kernelinfo
);
1669 if ((jobref
== -1) || (jobref
== 0))
1672 TAILQ_FOREACH(cb
, &ki
->kaio_jobdone
, plist
) {
1673 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1675 uap
->sysmsg_result
= cb
->uaiocb
._aiocb_private
.error
;
1682 for (cb
= TAILQ_FIRST(&ki
->kaio_jobqueue
); cb
; cb
= TAILQ_NEXT(cb
,
1684 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1686 uap
->sysmsg_result
= EINPROGRESS
;
1692 for (cb
= TAILQ_FIRST(&ki
->kaio_sockqueue
); cb
; cb
= TAILQ_NEXT(cb
,
1694 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1696 uap
->sysmsg_result
= EINPROGRESS
;
1704 for (cb
= TAILQ_FIRST(&ki
->kaio_bufdone
); cb
; cb
= TAILQ_NEXT(cb
,
1706 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1708 uap
->sysmsg_result
= cb
->uaiocb
._aiocb_private
.error
;
1714 for (cb
= TAILQ_FIRST(&ki
->kaio_bufqueue
); cb
; cb
= TAILQ_NEXT(cb
,
1716 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
) ==
1718 uap
->sysmsg_result
= EINPROGRESS
;
1729 status
= fuword(&uap
->aiocbp
->_aiocb_private
.status
);
1731 return fuword(&uap
->aiocbp
->_aiocb_private
.error
);
1734 #endif /* VFS_AIO */
1737 /* syscall - asynchronous read from a file (REALTIME) */
1739 sys_aio_read(struct aio_read_args
*uap
)
1744 return aio_aqueue(uap
->aiocbp
, LIO_READ
);
1745 #endif /* VFS_AIO */
1748 /* syscall - asynchronous write to a file (REALTIME) */
1750 sys_aio_write(struct aio_write_args
*uap
)
1755 return aio_aqueue(uap
->aiocbp
, LIO_WRITE
);
1756 #endif /* VFS_AIO */
1759 /* syscall - XXX undocumented */
1761 sys_lio_listio(struct lio_listio_args
*uap
)
1766 struct proc
*p
= curproc
;
1767 struct lwp
*lp
= curthread
->td_lwp
;
1768 int nent
, nentqueued
;
1769 struct aiocb
*iocb
, * const *cbptr
;
1770 struct aiocblist
*cb
;
1771 struct kaioinfo
*ki
;
1772 struct aio_liojob
*lj
;
1773 int error
, runningcode
;
1777 if ((uap
->mode
!= LIO_NOWAIT
) && (uap
->mode
!= LIO_WAIT
))
1781 if (nent
> AIO_LISTIO_MAX
)
1784 if (p
->p_aioinfo
== NULL
)
1785 aio_init_aioinfo(p
);
1787 if ((nent
+ num_queue_count
) > max_queue_count
)
1791 if ((nent
+ ki
->kaio_queue_count
) > ki
->kaio_qallowed_count
)
1794 lj
= zalloc(aiolio_zone
);
1799 lj
->lioj_buffer_count
= 0;
1800 lj
->lioj_buffer_finished_count
= 0;
1801 lj
->lioj_queue_count
= 0;
1802 lj
->lioj_queue_finished_count
= 0;
1808 if (uap
->sig
&& (uap
->mode
== LIO_NOWAIT
)) {
1809 error
= copyin(uap
->sig
, &lj
->lioj_signal
,
1810 sizeof(lj
->lioj_signal
));
1812 zfree(aiolio_zone
, lj
);
1815 if (!_SIG_VALID(lj
->lioj_signal
.sigev_signo
)) {
1816 zfree(aiolio_zone
, lj
);
1819 lj
->lioj_flags
|= LIOJ_SIGNAL
;
1820 lj
->lioj_flags
&= ~LIOJ_SIGNAL_POSTED
;
1822 lj
->lioj_flags
&= ~LIOJ_SIGNAL
;
1824 TAILQ_INSERT_TAIL(&ki
->kaio_liojoblist
, lj
, lioj_list
);
1826 * Get pointers to the list of I/O requests.
1830 cbptr
= uap
->acb_list
;
1831 for (i
= 0; i
< uap
->nent
; i
++) {
1832 iocb
= (struct aiocb
*)(intptr_t)fuword(&cbptr
[i
]);
1833 if (((intptr_t)iocb
!= -1) && ((intptr_t)iocb
!= 0)) {
1834 error
= _aio_aqueue(iocb
, lj
, 0);
1843 * If we haven't queued any, then just return error.
1845 if (nentqueued
== 0)
1849 * Calculate the appropriate error return.
1855 if (uap
->mode
== LIO_WAIT
) {
1856 int command
, found
, jobref
;
1860 for (i
= 0; i
< uap
->nent
; i
++) {
1862 * Fetch address of the control buf pointer in
1865 iocb
= (struct aiocb
*)
1866 (intptr_t)fuword(&cbptr
[i
]);
1867 if (((intptr_t)iocb
== -1) || ((intptr_t)iocb
1872 * Fetch the associated command from user space.
1874 command
= fuword(&iocb
->aio_lio_opcode
);
1875 if (command
== LIO_NOP
) {
1880 jobref
= fuword(&iocb
->_aiocb_private
.kernelinfo
);
1882 TAILQ_FOREACH(cb
, &ki
->kaio_jobdone
, plist
) {
1883 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
)
1885 if (cb
->uaiocb
.aio_lio_opcode
1887 lp
->lwp_ru
.ru_oublock
+=
1889 cb
->outputcharge
= 0;
1890 } else if (cb
->uaiocb
.aio_lio_opcode
1892 lp
->lwp_ru
.ru_inblock
+=
1894 cb
->inputcharge
= 0;
1902 TAILQ_FOREACH(cb
, &ki
->kaio_bufdone
, plist
) {
1903 if (((intptr_t)cb
->uaiocb
._aiocb_private
.kernelinfo
)
1913 * If all I/Os have been disposed of, then we can
1916 if (found
== nentqueued
)
1919 ki
->kaio_flags
|= KAIO_WAKEUP
;
1920 error
= tsleep(p
, PCATCH
, "aiospn", 0);
1924 else if (error
== EWOULDBLOCK
)
1930 #endif /* VFS_AIO */
1935 * This is a weird hack so that we can post a signal. It is safe to do so from
1936 * a timeout routine, but *not* from an interrupt routine.
1939 process_signal(void *aioj
)
1941 struct aiocblist
*aiocbe
= aioj
;
1942 struct aio_liojob
*lj
= aiocbe
->lio
;
1943 struct aiocb
*cb
= &aiocbe
->uaiocb
;
1945 if ((lj
) && (lj
->lioj_signal
.sigev_notify
== SIGEV_SIGNAL
) &&
1946 (lj
->lioj_queue_count
== lj
->lioj_queue_finished_count
)) {
1947 ksignal(lj
->lioj_ki
->kaio_p
, lj
->lioj_signal
.sigev_signo
);
1948 lj
->lioj_flags
|= LIOJ_SIGNAL_POSTED
;
1951 if (cb
->aio_sigevent
.sigev_notify
== SIGEV_SIGNAL
)
1952 ksignal(aiocbe
->userproc
, cb
->aio_sigevent
.sigev_signo
);
1956 * Interrupt handler for physio, performs the necessary process wakeups, and
1960 aio_physwakeup(struct bio
*bio
)
1962 struct buf
*bp
= bio
->bio_buf
;
1963 struct aiocblist
*aiocbe
;
1965 struct kaioinfo
*ki
;
1966 struct aio_liojob
*lj
;
1968 aiocbe
= bio
->bio_caller_info2
.ptr
;
1971 p
= bio
->bio_caller_info1
.ptr
;
1973 aiocbe
->jobstate
= JOBST_JOBBFINISHED
;
1974 aiocbe
->uaiocb
._aiocb_private
.status
-= bp
->b_resid
;
1975 aiocbe
->uaiocb
._aiocb_private
.error
= 0;
1976 aiocbe
->jobflags
|= AIOCBLIST_DONE
;
1978 if (bp
->b_flags
& B_ERROR
)
1979 aiocbe
->uaiocb
._aiocb_private
.error
= bp
->b_error
;
1983 lj
->lioj_buffer_finished_count
++;
1986 * wakeup/signal if all of the interrupt jobs are done.
1988 if (lj
->lioj_buffer_finished_count
==
1989 lj
->lioj_buffer_count
) {
1991 * Post a signal if it is called for.
1993 if ((lj
->lioj_flags
&
1994 (LIOJ_SIGNAL
|LIOJ_SIGNAL_POSTED
)) ==
1996 lj
->lioj_flags
|= LIOJ_SIGNAL_POSTED
;
1997 callout_reset(&aiocbe
->timeout
, 0,
1998 process_signal
, aiocbe
);
2005 ki
->kaio_buffer_finished_count
++;
2006 TAILQ_REMOVE(&aio_bufjobs
, aiocbe
, list
);
2007 TAILQ_REMOVE(&ki
->kaio_bufqueue
, aiocbe
, plist
);
2008 TAILQ_INSERT_TAIL(&ki
->kaio_bufdone
, aiocbe
, plist
);
2010 KNOTE(&aiocbe
->klist
, 0);
2011 /* Do the wakeup. */
2012 if (ki
->kaio_flags
& (KAIO_RUNDOWN
|KAIO_WAKEUP
)) {
2013 ki
->kaio_flags
&= ~KAIO_WAKEUP
;
2018 if (aiocbe
->uaiocb
.aio_sigevent
.sigev_notify
== SIGEV_SIGNAL
) {
2019 callout_reset(&aiocbe
->timeout
, 0,
2020 process_signal
, aiocbe
);
2025 #endif /* VFS_AIO */
2027 /* syscall - wait for the next completion of an aio request */
2029 sys_aio_waitcomplete(struct aio_waitcomplete_args
*uap
)
2034 struct proc
*p
= curproc
;
2035 struct lwp
*lp
= curthread
->td_lwp
;
2038 struct kaioinfo
*ki
;
2039 struct aiocblist
*cb
= NULL
;
2042 suword(uap
->aiocbp
, (int)NULL
);
2046 /* Get timespec struct. */
2047 error
= copyin(uap
->timeout
, &ts
, sizeof(ts
));
2051 if ((ts
.tv_nsec
< 0) || (ts
.tv_nsec
>= 1000000000))
2054 TIMESPEC_TO_TIMEVAL(&atv
, &ts
);
2055 if (itimerfix(&atv
))
2057 timo
= tvtohz_high(&atv
);
2065 if ((cb
= TAILQ_FIRST(&ki
->kaio_jobdone
)) != 0) {
2066 suword(uap
->aiocbp
, (uintptr_t)cb
->uuaiocb
);
2067 uap
->sysmsg_result
= cb
->uaiocb
._aiocb_private
.status
;
2068 if (cb
->uaiocb
.aio_lio_opcode
== LIO_WRITE
) {
2069 lp
->lwp_ru
.ru_oublock
+=
2071 cb
->outputcharge
= 0;
2072 } else if (cb
->uaiocb
.aio_lio_opcode
== LIO_READ
) {
2073 lp
->lwp_ru
.ru_inblock
+= cb
->inputcharge
;
2074 cb
->inputcharge
= 0;
2077 return cb
->uaiocb
._aiocb_private
.error
;
2081 if ((cb
= TAILQ_FIRST(&ki
->kaio_bufdone
)) != 0 ) {
2083 suword(uap
->aiocbp
, (uintptr_t)cb
->uuaiocb
);
2084 uap
->sysmsg_result
= cb
->uaiocb
._aiocb_private
.status
;
2086 return cb
->uaiocb
._aiocb_private
.error
;
2089 ki
->kaio_flags
|= KAIO_WAKEUP
;
2090 error
= tsleep(p
, PCATCH
, "aiowc", timo
);
2093 if (error
== ERESTART
)
2097 else if (error
== EINTR
)
2099 else if (error
== EWOULDBLOCK
)
2102 #endif /* VFS_AIO */
2107 filt_aioattach(struct knote
*kn
)
2113 struct filterops aio_filtops
=
2114 { 0, filt_aioattach
, NULL
, NULL
};
2117 /* kqueue attach function */
2119 filt_aioattach(struct knote
*kn
)
2121 struct aiocblist
*aiocbe
= (struct aiocblist
*)kn
->kn_sdata
;
2124 * The aiocbe pointer must be validated before using it, so
2125 * registration is restricted to the kernel; the user cannot
2128 if ((kn
->kn_flags
& EV_FLAG1
) == 0)
2130 kn
->kn_flags
&= ~EV_FLAG1
;
2132 SLIST_INSERT_HEAD(&aiocbe
->klist
, kn
, kn_selnext
);
2137 /* kqueue detach function */
2139 filt_aiodetach(struct knote
*kn
)
2141 struct aiocblist
*aiocbe
= (struct aiocblist
*)kn
->kn_sdata
;
2143 SLIST_REMOVE(&aiocbe
->klist
, kn
, knote
, kn_selnext
);
2146 /* kqueue filter function */
2149 filt_aio(struct knote
*kn
, long hint
)
2151 struct aiocblist
*aiocbe
= (struct aiocblist
*)kn
->kn_sdata
;
2153 kn
->kn_data
= aiocbe
->uaiocb
._aiocb_private
.error
;
2154 if (aiocbe
->jobstate
!= JOBST_JOBFINISHED
&&
2155 aiocbe
->jobstate
!= JOBST_JOBBFINISHED
)
2157 kn
->kn_flags
|= EV_EOF
;
2161 struct filterops aio_filtops
=
2162 { 0, filt_aioattach
, filt_aiodetach
, filt_aio
};
2163 #endif /* VFS_AIO */