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 1.40 2007/04/30 07:18:54 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 *curcp, *mycp, *userp;
623 struct vmspace *myvm;
627 * Local copies of curproc (cp) and vmspace (myvm). Get extra
628 * reference on myvm so we can use pmap_replacevm()
631 myvm = mycp->p_vmspace;
632 sysref_get(&myvm->vm_sysref);
634 if (mycp->p_textvp) {
635 vrele(mycp->p_textvp);
636 mycp->p_textvp = NULL;
640 * Allocate and ready the aio control info. There is one aiop structure
643 aiop = zalloc(aiop_zone);
644 aiop->aioproc = mycp;
645 aiop->aioprocflags |= AIOP_FREE;
650 * Place thread (lightweight process) onto the AIO free thread list.
652 if (TAILQ_EMPTY(&aio_freeproc))
653 wakeup(&aio_freeproc);
654 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
658 /* Make up a name for the daemon. */
659 strcpy(mycp->p_comm, "aiod");
662 * Get rid of our current filedescriptors. AIOD's don't need any
663 * filedescriptors, except as temporarily inherited from the client.
664 * Credentials are also cloned, and made equivalent to "root".
668 cr = cratom(&mycp->p_ucred);
670 uireplace(&cr->cr_uidinfo, uifind(0));
672 cr->cr_groups[0] = 1;
674 /* The daemon resides in its own pgrp. */
675 enterpgrp(mycp, mycp->p_pid, 1);
677 /* Mark special process type. */
678 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
681 * Wakeup parent process. (Parent sleeps to keep from blasting away
682 * and creating too many daemons.)
688 * curcp is the current daemon process context.
689 * userp is the current user process context.
694 * Take daemon off of free queue
696 if (aiop->aioprocflags & AIOP_FREE) {
698 TAILQ_REMOVE(&aio_freeproc, aiop, list);
699 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
700 aiop->aioprocflags &= ~AIOP_FREE;
703 aiop->aioprocflags &= ~AIOP_SCHED;
708 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
709 cb = &aiocbe->uaiocb;
710 userp = aiocbe->userproc;
712 aiocbe->jobstate = JOBST_JOBRUNNING;
715 * Connect to process address space for user program.
717 if (userp != curcp) {
718 pmap_replacevm(mycp, userp->p_vmspace, 1);
722 ki = userp->p_aioinfo;
725 /* Account for currently active jobs. */
726 ki->kaio_active_count++;
728 /* Do the I/O function. */
731 /* Decrement the active job count. */
732 ki->kaio_active_count--;
735 * Increment the completion count for wakeup/signal
738 aiocbe->jobflags |= AIOCBLIST_DONE;
739 ki->kaio_queue_finished_count++;
741 lj->lioj_queue_finished_count++;
742 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
743 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
744 ki->kaio_flags &= ~KAIO_WAKEUP;
749 if (lj && (lj->lioj_flags &
750 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
751 if ((lj->lioj_queue_finished_count ==
752 lj->lioj_queue_count) &&
753 (lj->lioj_buffer_finished_count ==
754 lj->lioj_buffer_count)) {
756 lj->lioj_signal.sigev_signo);
763 aiocbe->jobstate = JOBST_JOBFINISHED;
766 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
767 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
769 KNOTE(&aiocbe->klist, 0);
771 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
773 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
776 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
777 ksignal(userp, cb->aio_sigevent.sigev_signo);
782 * Disconnect from user address space.
785 /* swap our original address space back in */
786 pmap_replacevm(mycp, myvm, 1);
791 * If we are the first to be put onto the free queue, wakeup
792 * anyone waiting for a daemon.
795 TAILQ_REMOVE(&aio_activeproc, aiop, list);
796 if (TAILQ_EMPTY(&aio_freeproc))
797 wakeup(&aio_freeproc);
798 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
799 aiop->aioprocflags |= AIOP_FREE;
803 * If daemon is inactive for a long time, allow it to exit,
804 * thereby freeing resources.
806 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
807 0, "aiordy", aiod_lifetime)) {
809 if (TAILQ_EMPTY(&aio_jobs)) {
810 if ((aiop->aioprocflags & AIOP_FREE) &&
811 (num_aio_procs > target_aio_procs)) {
812 TAILQ_REMOVE(&aio_freeproc, aiop, list);
814 zfree(aiop_zone, aiop);
817 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
818 kprintf("AIOD: bad vm refcnt for"
819 " exiting daemon: %d\n",
820 mycp->p_vmspace->vm_sysref.refcnt);
823 sysref_put(&myvm->vm_sysref);
833 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
834 * AIO daemon modifies its environment itself.
840 struct lwp *lp, *nlp;
844 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
847 nlp = ONLY_LWP_IN_PROC(np);
848 cpu_set_fork_handler(nlp, aio_daemon, curproc);
849 start_forked_proc(lp, np);
852 * Wait until daemon is started, but continue on just in case to
853 * handle error conditions.
855 error = tsleep(np, 0, "aiosta", aiod_timeout);
862 * Try the high-performance, low-overhead physio method for eligible
863 * VCHR devices. This method doesn't use an aio helper thread, and
864 * thus has very low overhead.
866 * Assumes that the caller, _aio_aqueue(), has incremented the file
867 * structure's reference count, preventing its deallocation for the
868 * duration of this call.
871 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
879 struct aio_liojob *lj;
882 cb = &aiocbe->uaiocb;
883 fp = aiocbe->fd_file;
885 if (fp->f_type != DTYPE_VNODE)
888 vp = (struct vnode *)fp->f_data;
891 * If its not a disk, we don't want to return a positive error.
892 * It causes the aio code to not fall through to try the thread
893 * way when you're talking to a regular file.
895 if (!vn_isdisk(vp, &error)) {
896 if (error == ENOTBLK)
902 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
906 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
910 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
913 ki->kaio_buffer_count++;
917 lj->lioj_buffer_count++;
919 /* Create and build a buffer header for a transfer. */
924 * Get a copy of the kva from the physical buffer.
926 bp->b_bio1.bio_caller_info1.ptr = p;
929 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
930 BUF_CMD_WRITE : BUF_CMD_READ;
931 bp->b_bio1.bio_done = aio_physwakeup;
932 bp->b_bio1.bio_offset = cb->aio_offset;
934 /* Bring buffer into kernel space. */
935 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
943 bp->b_bio1.bio_caller_info2.ptr = aiocbe;
944 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
945 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
946 aiocbe->jobstate = JOBST_JOBQBUF;
947 cb->_aiocb_private.status = cb->aio_nbytes;
953 /* Perform transfer. */
954 dev_dstrategy(vp->v_rdev, &bp->b_bio1);
960 * If we had an error invoking the request, or an error in processing
961 * the request before we have returned, we process it as an error in
962 * transfer. Note that such an I/O error is not indicated immediately,
963 * but is returned using the aio_error mechanism. In this case,
964 * aio_suspend will return immediately.
966 if (bp->b_error || (bp->b_flags & B_ERROR)) {
967 struct aiocb *job = aiocbe->uuaiocb;
969 aiocbe->uaiocb._aiocb_private.status = 0;
970 suword(&job->_aiocb_private.status, 0);
971 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
972 suword(&job->_aiocb_private.error, bp->b_error);
974 ki->kaio_buffer_finished_count++;
976 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
977 aiocbe->jobstate = JOBST_JOBBFINISHED;
978 aiocbe->jobflags |= AIOCBLIST_DONE;
979 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
980 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
981 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)
1011 while (bp->b_cmd != BUF_CMD_DONE) {
1012 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
1013 if (bp->b_cmd != BUF_CMD_DONE) {
1023 /* Release mapping into kernel space. */
1029 /* Check for an error. */
1030 if (bp->b_flags & B_ERROR)
1031 error = bp->b_error;
1036 #endif /* VFS_AIO */
1039 * Wake up aio requests that may be serviceable now.
1042 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1047 struct aiocblist *cb,*cbn;
1049 struct kaioinfo *ki = NULL;
1050 int opcode, wakecount = 0;
1051 struct aioproclist *aiop;
1053 if (ssb == &so->so_snd) {
1055 so->so_snd.ssb_flags &= ~SSB_AIO;
1058 so->so_rcv.ssb_flags &= ~SSB_AIO;
1061 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1062 cbn = TAILQ_NEXT(cb, list);
1063 if (opcode == cb->uaiocb.aio_lio_opcode) {
1066 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1067 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1068 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1069 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1071 if (cb->jobstate != JOBST_JOBQGLOBAL)
1072 panic("invalid queue value");
1076 while (wakecount--) {
1077 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1078 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1079 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1080 aiop->aioprocflags &= ~AIOP_FREE;
1081 wakeup(aiop->aioproc);
1084 #endif /* VFS_AIO */
1089 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1090 * technique is done in this code.
1093 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1095 struct proc *p = curproc;
1096 struct filedesc *fdp;
1101 int opcode, user_opcode;
1102 struct aiocblist *aiocbe;
1103 struct aioproclist *aiop;
1104 struct kaioinfo *ki;
1109 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1110 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1112 aiocbe = zalloc (aiocb_zone);
1114 aiocbe->inputcharge = 0;
1115 aiocbe->outputcharge = 0;
1116 callout_init(&aiocbe->timeout);
1117 SLIST_INIT(&aiocbe->klist);
1119 suword(&job->_aiocb_private.status, -1);
1120 suword(&job->_aiocb_private.error, 0);
1121 suword(&job->_aiocb_private.kernelinfo, -1);
1123 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1125 suword(&job->_aiocb_private.error, error);
1126 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1129 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1130 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1131 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1135 /* Save userspace address of the job info. */
1136 aiocbe->uuaiocb = job;
1138 /* Get the opcode. */
1139 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1140 if (type != LIO_NOP)
1141 aiocbe->uaiocb.aio_lio_opcode = type;
1142 opcode = aiocbe->uaiocb.aio_lio_opcode;
1144 /* Get the fd info for process. */
1148 * Range check file descriptor.
1150 fd = aiocbe->uaiocb.aio_fildes;
1151 if (fd >= fdp->fd_nfiles) {
1152 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1154 suword(&job->_aiocb_private.error, EBADF);
1158 fp = aiocbe->fd_file = fdp->fd_files[fd].fp;
1159 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1161 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1163 suword(&job->_aiocb_private.error, EBADF);
1168 if (aiocbe->uaiocb.aio_offset == -1LL) {
1172 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1177 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1178 if (jobrefid == LONG_MAX)
1183 if (opcode == LIO_NOP) {
1185 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1187 suword(&job->_aiocb_private.error, 0);
1188 suword(&job->_aiocb_private.status, 0);
1189 suword(&job->_aiocb_private.kernelinfo, 0);
1193 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1195 suword(&job->_aiocb_private.status, 0);
1200 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1201 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1202 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1206 * This method for requesting kevent-based notification won't
1207 * work on the alpha, since we're passing in a pointer
1208 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1209 * based method instead.
1211 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1212 user_opcode == LIO_WRITE)
1215 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1220 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1221 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL ||
1222 (kq_fp->f_type != DTYPE_KQUEUE)) {
1226 kq = (struct kqueue *)kq_fp->f_data;
1227 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1228 kev.filter = EVFILT_AIO;
1229 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1230 kev.data = (intptr_t)aiocbe;
1231 /* XXX lwp kqueue_register takes a thread, but only uses its proc */
1232 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1236 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1238 suword(&job->_aiocb_private.error, error);
1243 suword(&job->_aiocb_private.error, EINPROGRESS);
1244 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1245 aiocbe->userproc = p;
1246 aiocbe->jobflags = 0;
1250 if (fp->f_type == DTYPE_SOCKET) {
1252 * Alternate queueing for socket ops: Reach down into the
1253 * descriptor to get the socket data. Then check to see if the
1254 * socket is ready to be read or written (based on the requested
1257 * If it is not ready for io, then queue the aiocbe on the
1258 * socket, and set the flags so we get a call when ssb_notify()
1261 so = (struct socket *)fp->f_data;
1263 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1264 LIO_WRITE) && (!sowriteable(so)))) {
1265 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1266 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1267 if (opcode == LIO_READ)
1268 so->so_rcv.ssb_flags |= SSB_AIO;
1270 so->so_snd.ssb_flags |= SSB_AIO;
1271 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1272 ki->kaio_queue_count++;
1281 if ((error = aio_qphysio(p, aiocbe)) == 0)
1284 suword(&job->_aiocb_private.status, 0);
1285 aiocbe->uaiocb._aiocb_private.error = error;
1286 suword(&job->_aiocb_private.error, error);
1290 /* No buffer for daemon I/O. */
1293 ki->kaio_queue_count++;
1295 lj->lioj_queue_count++;
1297 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1298 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1300 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1306 * If we don't have a free AIO process, and we are below our quota, then
1307 * start one. Otherwise, depend on the subsequent I/O completions to
1308 * pick-up this job. If we don't successfully create the new process
1309 * (thread) due to resource issues, we return an error for now (EAGAIN),
1310 * which is likely not the correct thing to do.
1314 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1315 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1316 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1317 aiop->aioprocflags &= ~AIOP_FREE;
1318 wakeup(aiop->aioproc);
1319 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1320 ((ki->kaio_active_count + num_aio_resv_start) <
1321 ki->kaio_maxactive_count)) {
1322 num_aio_resv_start++;
1323 if ((error = aio_newproc()) == 0) {
1324 num_aio_resv_start--;
1327 num_aio_resv_start--;
1335 * This routine queues an AIO request, checking for quotas.
1338 aio_aqueue(struct aiocb *job, int type)
1340 struct proc *p = curproc;
1341 struct kaioinfo *ki;
1343 if (p->p_aioinfo == NULL)
1344 aio_init_aioinfo(p);
1346 if (num_queue_count >= max_queue_count)
1350 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1353 return _aio_aqueue(job, NULL, type);
1355 #endif /* VFS_AIO */
1358 * Support the aio_return system call, as a side-effect, kernel resources are
1362 sys_aio_return(struct aio_return_args *uap)
1367 struct proc *p = curproc;
1368 struct lwp *lp = curthread->td_lwp;
1370 struct aiocblist *cb, *ncb;
1372 struct kaioinfo *ki;
1380 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1381 if (jobref == -1 || jobref == 0)
1384 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1385 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1387 if (ujob == cb->uuaiocb) {
1388 uap->sysmsg_result =
1389 cb->uaiocb._aiocb_private.status;
1391 uap->sysmsg_result = EFAULT;
1392 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1393 lp->lwp_ru.ru_oublock += cb->outputcharge;
1394 cb->outputcharge = 0;
1395 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1396 lp->lwp_ru.ru_inblock += cb->inputcharge;
1397 cb->inputcharge = 0;
1404 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1405 ncb = TAILQ_NEXT(cb, plist);
1406 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1409 if (ujob == cb->uuaiocb) {
1410 uap->sysmsg_result =
1411 cb->uaiocb._aiocb_private.status;
1413 uap->sysmsg_result = EFAULT;
1421 #endif /* VFS_AIO */
1425 * Allow a process to wakeup when any of the I/O requests are completed.
1428 sys_aio_suspend(struct aio_suspend_args *uap)
1433 struct proc *p = curproc;
1436 struct aiocb *const *cbptr, *cbp;
1437 struct kaioinfo *ki;
1438 struct aiocblist *cb;
1443 struct aiocb **ujoblist;
1445 if (uap->nent > AIO_LISTIO_MAX)
1450 /* Get timespec struct. */
1451 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1454 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1457 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1458 if (itimerfix(&atv))
1460 timo = tvtohz_high(&atv);
1468 ijoblist = zalloc(aiol_zone);
1469 ujoblist = zalloc(aiol_zone);
1470 cbptr = uap->aiocbp;
1472 for (i = 0; i < uap->nent; i++) {
1473 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1476 ujoblist[njoblist] = cbp;
1477 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1481 if (njoblist == 0) {
1482 zfree(aiol_zone, ijoblist);
1483 zfree(aiol_zone, ujoblist);
1489 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1490 for (i = 0; i < njoblist; i++) {
1492 cb->uaiocb._aiocb_private.kernelinfo) ==
1494 if (ujoblist[i] != cb->uuaiocb)
1496 zfree(aiol_zone, ijoblist);
1497 zfree(aiol_zone, ujoblist);
1504 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1505 TAILQ_NEXT(cb, plist)) {
1506 for (i = 0; i < njoblist; i++) {
1508 cb->uaiocb._aiocb_private.kernelinfo) ==
1511 if (ujoblist[i] != cb->uuaiocb)
1513 zfree(aiol_zone, ijoblist);
1514 zfree(aiol_zone, ujoblist);
1520 ki->kaio_flags |= KAIO_WAKEUP;
1521 error = tsleep(p, PCATCH, "aiospn", timo);
1524 if (error == ERESTART || error == EINTR) {
1525 zfree(aiol_zone, ijoblist);
1526 zfree(aiol_zone, ujoblist);
1528 } else if (error == EWOULDBLOCK) {
1529 zfree(aiol_zone, ijoblist);
1530 zfree(aiol_zone, ujoblist);
1537 #endif /* VFS_AIO */
1541 * aio_cancel cancels any non-physio aio operations not currently in
1545 sys_aio_cancel(struct aio_cancel_args *uap)
1550 struct proc *p = curproc;
1551 struct kaioinfo *ki;
1552 struct aiocblist *cbe, *cbn;
1554 struct filedesc *fdp;
1563 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1564 (fp = fdp->fd_files[uap->fd].fp) == NULL)
1567 if (fp->f_type == DTYPE_VNODE) {
1568 vp = (struct vnode *)fp->f_data;
1570 if (vn_isdisk(vp,&error)) {
1571 uap->sysmsg_result = AIO_NOTCANCELED;
1574 } else if (fp->f_type == DTYPE_SOCKET) {
1575 so = (struct socket *)fp->f_data;
1579 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1580 cbn = TAILQ_NEXT(cbe, list);
1581 if ((uap->aiocbp == NULL) ||
1582 (uap->aiocbp == cbe->uuaiocb) ) {
1585 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1586 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1587 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1588 if (ki->kaio_flags & KAIO_WAKEUP) {
1591 cbe->jobstate = JOBST_JOBFINISHED;
1592 cbe->uaiocb._aiocb_private.status=-1;
1593 cbe->uaiocb._aiocb_private.error=ECANCELED;
1595 /* XXX cancelled, knote? */
1596 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1598 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1605 if ((cancelled) && (uap->aiocbp)) {
1606 uap->sysmsg_result = AIO_CANCELED;
1615 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1616 cbn = TAILQ_NEXT(cbe, plist);
1618 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1619 ((uap->aiocbp == NULL ) ||
1620 (uap->aiocbp == cbe->uuaiocb))) {
1622 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1623 TAILQ_REMOVE(&aio_jobs, cbe, list);
1624 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1625 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1628 ki->kaio_queue_finished_count++;
1629 cbe->jobstate = JOBST_JOBFINISHED;
1630 cbe->uaiocb._aiocb_private.status = -1;
1631 cbe->uaiocb._aiocb_private.error = ECANCELED;
1632 /* XXX cancelled, knote? */
1633 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1635 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1644 uap->sysmsg_result = AIO_NOTCANCELED;
1648 uap->sysmsg_result = AIO_CANCELED;
1651 uap->sysmsg_result = AIO_ALLDONE;
1654 #endif /* VFS_AIO */
1658 * aio_error is implemented in the kernel level for compatibility purposes only.
1659 * For a user mode async implementation, it would be best to do it in a userland
1663 sys_aio_error(struct aio_error_args *uap)
1668 struct proc *p = curproc;
1669 struct aiocblist *cb;
1670 struct kaioinfo *ki;
1677 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1678 if ((jobref == -1) || (jobref == 0))
1681 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1682 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1684 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1691 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1693 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1695 uap->sysmsg_result = EINPROGRESS;
1701 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1703 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1705 uap->sysmsg_result = EINPROGRESS;
1713 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1715 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1717 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1723 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1725 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1727 uap->sysmsg_result = EINPROGRESS;
1738 status = fuword(&uap->aiocbp->_aiocb_private.status);
1740 return fuword(&uap->aiocbp->_aiocb_private.error);
1743 #endif /* VFS_AIO */
1746 /* syscall - asynchronous read from a file (REALTIME) */
1748 sys_aio_read(struct aio_read_args *uap)
1753 return aio_aqueue(uap->aiocbp, LIO_READ);
1754 #endif /* VFS_AIO */
1757 /* syscall - asynchronous write to a file (REALTIME) */
1759 sys_aio_write(struct aio_write_args *uap)
1764 return aio_aqueue(uap->aiocbp, LIO_WRITE);
1765 #endif /* VFS_AIO */
1768 /* syscall - XXX undocumented */
1770 sys_lio_listio(struct lio_listio_args *uap)
1775 struct proc *p = curproc;
1776 struct lwp *lp = curthread->td_lwp;
1777 int nent, nentqueued;
1778 struct aiocb *iocb, * const *cbptr;
1779 struct aiocblist *cb;
1780 struct kaioinfo *ki;
1781 struct aio_liojob *lj;
1782 int error, runningcode;
1786 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1790 if (nent > AIO_LISTIO_MAX)
1793 if (p->p_aioinfo == NULL)
1794 aio_init_aioinfo(p);
1796 if ((nent + num_queue_count) > max_queue_count)
1800 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1803 lj = zalloc(aiolio_zone);
1808 lj->lioj_buffer_count = 0;
1809 lj->lioj_buffer_finished_count = 0;
1810 lj->lioj_queue_count = 0;
1811 lj->lioj_queue_finished_count = 0;
1817 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1818 error = copyin(uap->sig, &lj->lioj_signal,
1819 sizeof(lj->lioj_signal));
1821 zfree(aiolio_zone, lj);
1824 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1825 zfree(aiolio_zone, lj);
1828 lj->lioj_flags |= LIOJ_SIGNAL;
1829 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1831 lj->lioj_flags &= ~LIOJ_SIGNAL;
1833 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1835 * Get pointers to the list of I/O requests.
1839 cbptr = uap->acb_list;
1840 for (i = 0; i < uap->nent; i++) {
1841 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1842 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1843 error = _aio_aqueue(iocb, lj, 0);
1852 * If we haven't queued any, then just return error.
1854 if (nentqueued == 0)
1858 * Calculate the appropriate error return.
1864 if (uap->mode == LIO_WAIT) {
1865 int command, found, jobref;
1869 for (i = 0; i < uap->nent; i++) {
1871 * Fetch address of the control buf pointer in
1874 iocb = (struct aiocb *)
1875 (intptr_t)fuword(&cbptr[i]);
1876 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1881 * Fetch the associated command from user space.
1883 command = fuword(&iocb->aio_lio_opcode);
1884 if (command == LIO_NOP) {
1889 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1891 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1892 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1894 if (cb->uaiocb.aio_lio_opcode
1896 lp->lwp_ru.ru_oublock +=
1898 cb->outputcharge = 0;
1899 } else if (cb->uaiocb.aio_lio_opcode
1901 lp->lwp_ru.ru_inblock +=
1903 cb->inputcharge = 0;
1911 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1912 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1922 * If all I/Os have been disposed of, then we can
1925 if (found == nentqueued)
1928 ki->kaio_flags |= KAIO_WAKEUP;
1929 error = tsleep(p, PCATCH, "aiospn", 0);
1933 else if (error == EWOULDBLOCK)
1939 #endif /* VFS_AIO */
1944 * This is a weird hack so that we can post a signal. It is safe to do so from
1945 * a timeout routine, but *not* from an interrupt routine.
1948 process_signal(void *aioj)
1950 struct aiocblist *aiocbe = aioj;
1951 struct aio_liojob *lj = aiocbe->lio;
1952 struct aiocb *cb = &aiocbe->uaiocb;
1954 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1955 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1956 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1957 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1960 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1961 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
1965 * Interrupt handler for physio, performs the necessary process wakeups, and
1969 aio_physwakeup(struct bio *bio)
1971 struct buf *bp = bio->bio_buf;
1972 struct aiocblist *aiocbe;
1974 struct kaioinfo *ki;
1975 struct aio_liojob *lj;
1977 aiocbe = bio->bio_caller_info2.ptr;
1980 p = bio->bio_caller_info1.ptr;
1982 aiocbe->jobstate = JOBST_JOBBFINISHED;
1983 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
1984 aiocbe->uaiocb._aiocb_private.error = 0;
1985 aiocbe->jobflags |= AIOCBLIST_DONE;
1987 if (bp->b_flags & B_ERROR)
1988 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1992 lj->lioj_buffer_finished_count++;
1995 * wakeup/signal if all of the interrupt jobs are done.
1997 if (lj->lioj_buffer_finished_count ==
1998 lj->lioj_buffer_count) {
2000 * Post a signal if it is called for.
2002 if ((lj->lioj_flags &
2003 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2005 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2006 callout_reset(&aiocbe->timeout, 0,
2007 process_signal, aiocbe);
2014 ki->kaio_buffer_finished_count++;
2015 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2016 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2017 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2019 KNOTE(&aiocbe->klist, 0);
2020 /* Do the wakeup. */
2021 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2022 ki->kaio_flags &= ~KAIO_WAKEUP;
2027 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2028 callout_reset(&aiocbe->timeout, 0,
2029 process_signal, aiocbe);
2032 bp->b_cmd = BUF_CMD_DONE;
2035 #endif /* VFS_AIO */
2037 /* syscall - wait for the next completion of an aio request */
2039 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2044 struct proc *p = curproc;
2045 struct lwp *lp = curthread->td_lwp;
2048 struct kaioinfo *ki;
2049 struct aiocblist *cb = NULL;
2052 suword(uap->aiocbp, (int)NULL);
2056 /* Get timespec struct. */
2057 error = copyin(uap->timeout, &ts, sizeof(ts));
2061 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2064 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2065 if (itimerfix(&atv))
2067 timo = tvtohz_high(&atv);
2075 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2076 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2077 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2078 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2079 lp->lwp_ru.ru_oublock +=
2081 cb->outputcharge = 0;
2082 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2083 lp->lwp_ru.ru_inblock += cb->inputcharge;
2084 cb->inputcharge = 0;
2087 return cb->uaiocb._aiocb_private.error;
2091 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2093 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2094 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2096 return cb->uaiocb._aiocb_private.error;
2099 ki->kaio_flags |= KAIO_WAKEUP;
2100 error = tsleep(p, PCATCH, "aiowc", timo);
2103 if (error == ERESTART)
2107 else if (error == EINTR)
2109 else if (error == EWOULDBLOCK)
2112 #endif /* VFS_AIO */
2117 filt_aioattach(struct knote *kn)
2123 struct filterops aio_filtops =
2124 { 0, filt_aioattach, NULL, NULL };
2127 /* kqueue attach function */
2129 filt_aioattach(struct knote *kn)
2131 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2134 * The aiocbe pointer must be validated before using it, so
2135 * registration is restricted to the kernel; the user cannot
2138 if ((kn->kn_flags & EV_FLAG1) == 0)
2140 kn->kn_flags &= ~EV_FLAG1;
2142 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2147 /* kqueue detach function */
2149 filt_aiodetach(struct knote *kn)
2151 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2153 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2156 /* kqueue filter function */
2159 filt_aio(struct knote *kn, long hint)
2161 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2163 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2164 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2165 aiocbe->jobstate != JOBST_JOBBFINISHED)
2167 kn->kn_flags |= EV_EOF;
2171 struct filterops aio_filtops =
2172 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2173 #endif /* VFS_AIO */