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".
664 cr = cratom(&mycp->p_ucred);
666 uireplace(&cr->cr_uidinfo, uifind(0));
668 cr->cr_groups[0] = 1;
670 /* The daemon resides in its own pgrp. */
671 enterpgrp(mycp, mycp->p_pid, 1);
673 /* Mark special process type. */
674 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
677 * Wakeup parent process. (Parent sleeps to keep from blasting away
678 * and creating too many daemons.)
685 * Take daemon off of free queue
687 if (aiop->aioprocflags & AIOP_FREE) {
689 TAILQ_REMOVE(&aio_freeproc, aiop, list);
690 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
691 aiop->aioprocflags &= ~AIOP_FREE;
694 aiop->aioprocflags &= ~AIOP_SCHED;
699 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
700 cb = &aiocbe->uaiocb;
701 userp = aiocbe->userproc;
703 aiocbe->jobstate = JOBST_JOBRUNNING;
706 * Connect to process address space for user program.
708 if (curvm != userp->p_vmspace) {
709 pmap_setlwpvm(mylwp, userp->p_vmspace);
711 sysref_put(&curvm->vm_sysref);
712 curvm = userp->p_vmspace;
713 sysref_get(&curvm->vm_sysref);
716 ki = userp->p_aioinfo;
719 /* Account for currently active jobs. */
720 ki->kaio_active_count++;
722 /* Do the I/O function. */
725 /* Decrement the active job count. */
726 ki->kaio_active_count--;
729 * Increment the completion count for wakeup/signal
732 aiocbe->jobflags |= AIOCBLIST_DONE;
733 ki->kaio_queue_finished_count++;
735 lj->lioj_queue_finished_count++;
736 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
737 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
738 ki->kaio_flags &= ~KAIO_WAKEUP;
743 if (lj && (lj->lioj_flags &
744 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
745 if ((lj->lioj_queue_finished_count ==
746 lj->lioj_queue_count) &&
747 (lj->lioj_buffer_finished_count ==
748 lj->lioj_buffer_count)) {
750 lj->lioj_signal.sigev_signo);
757 aiocbe->jobstate = JOBST_JOBFINISHED;
760 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
761 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
763 KNOTE(&aiocbe->klist, 0);
765 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
767 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
770 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
771 ksignal(userp, cb->aio_sigevent.sigev_signo);
776 * Disconnect from user address space.
779 /* swap our original address space back in */
780 pmap_setlwpvm(mylwp, mycp->p_vmspace);
781 sysref_put(&curvm->vm_sysref);
786 * If we are the first to be put onto the free queue, wakeup
787 * anyone waiting for a daemon.
790 TAILQ_REMOVE(&aio_activeproc, aiop, list);
791 if (TAILQ_EMPTY(&aio_freeproc))
792 wakeup(&aio_freeproc);
793 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
794 aiop->aioprocflags |= AIOP_FREE;
798 * If daemon is inactive for a long time, allow it to exit,
799 * thereby freeing resources.
801 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
802 0, "aiordy", aiod_lifetime)) {
804 if (TAILQ_EMPTY(&aio_jobs)) {
805 if ((aiop->aioprocflags & AIOP_FREE) &&
806 (num_aio_procs > target_aio_procs)) {
807 TAILQ_REMOVE(&aio_freeproc, aiop, list);
809 zfree(aiop_zone, aiop);
812 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
813 kprintf("AIOD: bad vm refcnt for"
814 " exiting daemon: %d\n",
815 mycp->p_vmspace->vm_sysref.refcnt);
827 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
828 * AIO daemon modifies its environment itself.
834 struct lwp *lp, *nlp;
838 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
841 nlp = ONLY_LWP_IN_PROC(np);
842 cpu_set_fork_handler(nlp, aio_daemon, curproc);
843 start_forked_proc(lp, np);
846 * Wait until daemon is started, but continue on just in case to
847 * handle error conditions.
849 error = tsleep(np, 0, "aiosta", aiod_timeout);
856 * Try the high-performance, low-overhead physio method for eligible
857 * VCHR devices. This method doesn't use an aio helper thread, and
858 * thus has very low overhead.
860 * Assumes that the caller, _aio_aqueue(), has incremented the file
861 * structure's reference count, preventing its deallocation for the
862 * duration of this call.
865 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
873 struct aio_liojob *lj;
876 cb = &aiocbe->uaiocb;
877 fp = aiocbe->fd_file;
879 if (fp->f_type != DTYPE_VNODE)
882 vp = (struct vnode *)fp->f_data;
885 * If its not a disk, we don't want to return a positive error.
886 * It causes the aio code to not fall through to try the thread
887 * way when you're talking to a regular file.
889 if (!vn_isdisk(vp, &error)) {
890 if (error == ENOTBLK)
896 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
900 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
904 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
907 ki->kaio_buffer_count++;
911 lj->lioj_buffer_count++;
913 /* Create and build a buffer header for a transfer. */
918 * Get a copy of the kva from the physical buffer.
920 bp->b_bio1.bio_caller_info1.ptr = p;
923 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
924 BUF_CMD_WRITE : BUF_CMD_READ;
925 bp->b_bio1.bio_done = aio_physwakeup;
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);
958 * If we had an error invoking the request, or an error in processing
959 * the request before we have returned, we process it as an error in
960 * transfer. Note that such an I/O error is not indicated immediately,
961 * but is returned using the aio_error mechanism. In this case,
962 * aio_suspend will return immediately.
964 if (bp->b_error || (bp->b_flags & B_ERROR)) {
965 struct aiocb *job = aiocbe->uuaiocb;
967 aiocbe->uaiocb._aiocb_private.status = 0;
968 suword(&job->_aiocb_private.status, 0);
969 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
970 suword(&job->_aiocb_private.error, bp->b_error);
972 ki->kaio_buffer_finished_count++;
974 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
975 aiocbe->jobstate = JOBST_JOBBFINISHED;
976 aiocbe->jobflags |= AIOCBLIST_DONE;
977 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
978 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
979 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
985 KNOTE(&aiocbe->klist, 0);
989 ki->kaio_buffer_count--;
991 lj->lioj_buffer_count--;
998 * This waits/tests physio completion.
1001 aio_fphysio(struct aiocblist *iocb)
1009 while (bp->b_cmd != BUF_CMD_DONE) {
1010 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
1011 if (bp->b_cmd != BUF_CMD_DONE) {
1021 /* Release mapping into kernel space. */
1027 /* Check for an error. */
1028 if (bp->b_flags & B_ERROR)
1029 error = bp->b_error;
1034 #endif /* VFS_AIO */
1037 * Wake up aio requests that may be serviceable now.
1040 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1045 struct aiocblist *cb,*cbn;
1047 struct kaioinfo *ki = NULL;
1048 int opcode, wakecount = 0;
1049 struct aioproclist *aiop;
1051 if (ssb == &so->so_snd) {
1053 so->so_snd.ssb_flags &= ~SSB_AIO;
1056 so->so_rcv.ssb_flags &= ~SSB_AIO;
1059 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1060 cbn = TAILQ_NEXT(cb, list);
1061 if (opcode == cb->uaiocb.aio_lio_opcode) {
1064 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1065 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1066 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1067 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1069 if (cb->jobstate != JOBST_JOBQGLOBAL)
1070 panic("invalid queue value");
1074 while (wakecount--) {
1075 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1076 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1077 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1078 aiop->aioprocflags &= ~AIOP_FREE;
1079 wakeup(aiop->aioproc);
1082 #endif /* VFS_AIO */
1087 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1088 * technique is done in this code.
1091 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1093 struct proc *p = curproc;
1094 struct filedesc *fdp;
1099 int opcode, user_opcode;
1100 struct aiocblist *aiocbe;
1101 struct aioproclist *aiop;
1102 struct kaioinfo *ki;
1107 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1108 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1110 aiocbe = zalloc (aiocb_zone);
1112 aiocbe->inputcharge = 0;
1113 aiocbe->outputcharge = 0;
1114 callout_init(&aiocbe->timeout);
1115 SLIST_INIT(&aiocbe->klist);
1117 suword(&job->_aiocb_private.status, -1);
1118 suword(&job->_aiocb_private.error, 0);
1119 suword(&job->_aiocb_private.kernelinfo, -1);
1121 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1123 suword(&job->_aiocb_private.error, error);
1124 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1127 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1128 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1129 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1133 /* Save userspace address of the job info. */
1134 aiocbe->uuaiocb = job;
1136 /* Get the opcode. */
1137 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1138 if (type != LIO_NOP)
1139 aiocbe->uaiocb.aio_lio_opcode = type;
1140 opcode = aiocbe->uaiocb.aio_lio_opcode;
1142 /* Get the fd info for process. */
1146 * Range check file descriptor.
1148 fd = aiocbe->uaiocb.aio_fildes;
1149 if (fd >= fdp->fd_nfiles) {
1150 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1152 suword(&job->_aiocb_private.error, EBADF);
1156 fp = aiocbe->fd_file = fdp->fd_files[fd].fp;
1157 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1159 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1161 suword(&job->_aiocb_private.error, EBADF);
1166 if (aiocbe->uaiocb.aio_offset == -1LL) {
1170 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1175 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1176 if (jobrefid == LONG_MAX)
1181 if (opcode == LIO_NOP) {
1183 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1185 suword(&job->_aiocb_private.error, 0);
1186 suword(&job->_aiocb_private.status, 0);
1187 suword(&job->_aiocb_private.kernelinfo, 0);
1191 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1193 suword(&job->_aiocb_private.status, 0);
1198 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1199 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1200 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1204 * This method for requesting kevent-based notification won't
1205 * work on the alpha, since we're passing in a pointer
1206 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1207 * based method instead.
1209 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1210 user_opcode == LIO_WRITE)
1213 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1218 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1219 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL ||
1220 (kq_fp->f_type != DTYPE_KQUEUE)) {
1224 kq = (struct kqueue *)kq_fp->f_data;
1225 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1226 kev.filter = EVFILT_AIO;
1227 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1228 kev.data = (intptr_t)aiocbe;
1229 /* XXX lwp kqueue_register takes a thread, but only uses its proc */
1230 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1234 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1236 suword(&job->_aiocb_private.error, error);
1241 suword(&job->_aiocb_private.error, EINPROGRESS);
1242 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1243 aiocbe->userproc = p;
1244 aiocbe->jobflags = 0;
1248 if (fp->f_type == DTYPE_SOCKET) {
1250 * Alternate queueing for socket ops: Reach down into the
1251 * descriptor to get the socket data. Then check to see if the
1252 * socket is ready to be read or written (based on the requested
1255 * If it is not ready for io, then queue the aiocbe on the
1256 * socket, and set the flags so we get a call when ssb_notify()
1259 so = (struct socket *)fp->f_data;
1261 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1262 LIO_WRITE) && (!sowriteable(so)))) {
1263 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1264 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1265 if (opcode == LIO_READ)
1266 so->so_rcv.ssb_flags |= SSB_AIO;
1268 so->so_snd.ssb_flags |= SSB_AIO;
1269 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1270 ki->kaio_queue_count++;
1279 if ((error = aio_qphysio(p, aiocbe)) == 0)
1282 suword(&job->_aiocb_private.status, 0);
1283 aiocbe->uaiocb._aiocb_private.error = error;
1284 suword(&job->_aiocb_private.error, error);
1288 /* No buffer for daemon I/O. */
1291 ki->kaio_queue_count++;
1293 lj->lioj_queue_count++;
1295 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1296 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1298 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1304 * If we don't have a free AIO process, and we are below our quota, then
1305 * start one. Otherwise, depend on the subsequent I/O completions to
1306 * pick-up this job. If we don't successfully create the new process
1307 * (thread) due to resource issues, we return an error for now (EAGAIN),
1308 * which is likely not the correct thing to do.
1312 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1313 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1314 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1315 aiop->aioprocflags &= ~AIOP_FREE;
1316 wakeup(aiop->aioproc);
1317 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1318 ((ki->kaio_active_count + num_aio_resv_start) <
1319 ki->kaio_maxactive_count)) {
1320 num_aio_resv_start++;
1321 if ((error = aio_newproc()) == 0) {
1322 num_aio_resv_start--;
1325 num_aio_resv_start--;
1333 * This routine queues an AIO request, checking for quotas.
1336 aio_aqueue(struct aiocb *job, int type)
1338 struct proc *p = curproc;
1339 struct kaioinfo *ki;
1341 if (p->p_aioinfo == NULL)
1342 aio_init_aioinfo(p);
1344 if (num_queue_count >= max_queue_count)
1348 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1351 return _aio_aqueue(job, NULL, type);
1353 #endif /* VFS_AIO */
1356 * Support the aio_return system call, as a side-effect, kernel resources are
1360 sys_aio_return(struct aio_return_args *uap)
1365 struct proc *p = curproc;
1366 struct lwp *lp = curthread->td_lwp;
1368 struct aiocblist *cb, *ncb;
1370 struct kaioinfo *ki;
1378 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1379 if (jobref == -1 || jobref == 0)
1382 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1383 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1385 if (ujob == cb->uuaiocb) {
1386 uap->sysmsg_result =
1387 cb->uaiocb._aiocb_private.status;
1389 uap->sysmsg_result = EFAULT;
1390 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1391 lp->lwp_ru.ru_oublock += cb->outputcharge;
1392 cb->outputcharge = 0;
1393 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1394 lp->lwp_ru.ru_inblock += cb->inputcharge;
1395 cb->inputcharge = 0;
1402 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1403 ncb = TAILQ_NEXT(cb, plist);
1404 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1407 if (ujob == cb->uuaiocb) {
1408 uap->sysmsg_result =
1409 cb->uaiocb._aiocb_private.status;
1411 uap->sysmsg_result = EFAULT;
1419 #endif /* VFS_AIO */
1423 * Allow a process to wakeup when any of the I/O requests are completed.
1426 sys_aio_suspend(struct aio_suspend_args *uap)
1431 struct proc *p = curproc;
1434 struct aiocb *const *cbptr, *cbp;
1435 struct kaioinfo *ki;
1436 struct aiocblist *cb;
1441 struct aiocb **ujoblist;
1443 if (uap->nent > AIO_LISTIO_MAX)
1448 /* Get timespec struct. */
1449 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1452 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1455 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1456 if (itimerfix(&atv))
1458 timo = tvtohz_high(&atv);
1466 ijoblist = zalloc(aiol_zone);
1467 ujoblist = zalloc(aiol_zone);
1468 cbptr = uap->aiocbp;
1470 for (i = 0; i < uap->nent; i++) {
1471 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1474 ujoblist[njoblist] = cbp;
1475 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1479 if (njoblist == 0) {
1480 zfree(aiol_zone, ijoblist);
1481 zfree(aiol_zone, ujoblist);
1487 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1488 for (i = 0; i < njoblist; i++) {
1490 cb->uaiocb._aiocb_private.kernelinfo) ==
1492 if (ujoblist[i] != cb->uuaiocb)
1494 zfree(aiol_zone, ijoblist);
1495 zfree(aiol_zone, ujoblist);
1502 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1503 TAILQ_NEXT(cb, plist)) {
1504 for (i = 0; i < njoblist; i++) {
1506 cb->uaiocb._aiocb_private.kernelinfo) ==
1509 if (ujoblist[i] != cb->uuaiocb)
1511 zfree(aiol_zone, ijoblist);
1512 zfree(aiol_zone, ujoblist);
1518 ki->kaio_flags |= KAIO_WAKEUP;
1519 error = tsleep(p, PCATCH, "aiospn", timo);
1522 if (error == ERESTART || error == EINTR) {
1523 zfree(aiol_zone, ijoblist);
1524 zfree(aiol_zone, ujoblist);
1526 } else if (error == EWOULDBLOCK) {
1527 zfree(aiol_zone, ijoblist);
1528 zfree(aiol_zone, ujoblist);
1535 #endif /* VFS_AIO */
1539 * aio_cancel cancels any non-physio aio operations not currently in
1543 sys_aio_cancel(struct aio_cancel_args *uap)
1548 struct proc *p = curproc;
1549 struct kaioinfo *ki;
1550 struct aiocblist *cbe, *cbn;
1552 struct filedesc *fdp;
1561 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1562 (fp = fdp->fd_files[uap->fd].fp) == NULL)
1565 if (fp->f_type == DTYPE_VNODE) {
1566 vp = (struct vnode *)fp->f_data;
1568 if (vn_isdisk(vp,&error)) {
1569 uap->sysmsg_result = AIO_NOTCANCELED;
1572 } else if (fp->f_type == DTYPE_SOCKET) {
1573 so = (struct socket *)fp->f_data;
1577 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1578 cbn = TAILQ_NEXT(cbe, list);
1579 if ((uap->aiocbp == NULL) ||
1580 (uap->aiocbp == cbe->uuaiocb) ) {
1583 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1584 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1585 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1586 if (ki->kaio_flags & KAIO_WAKEUP) {
1589 cbe->jobstate = JOBST_JOBFINISHED;
1590 cbe->uaiocb._aiocb_private.status=-1;
1591 cbe->uaiocb._aiocb_private.error=ECANCELED;
1593 /* XXX cancelled, knote? */
1594 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1596 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1603 if ((cancelled) && (uap->aiocbp)) {
1604 uap->sysmsg_result = AIO_CANCELED;
1613 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1614 cbn = TAILQ_NEXT(cbe, plist);
1616 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1617 ((uap->aiocbp == NULL ) ||
1618 (uap->aiocbp == cbe->uuaiocb))) {
1620 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1621 TAILQ_REMOVE(&aio_jobs, cbe, list);
1622 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1623 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1626 ki->kaio_queue_finished_count++;
1627 cbe->jobstate = JOBST_JOBFINISHED;
1628 cbe->uaiocb._aiocb_private.status = -1;
1629 cbe->uaiocb._aiocb_private.error = ECANCELED;
1630 /* XXX cancelled, knote? */
1631 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1633 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1642 uap->sysmsg_result = AIO_NOTCANCELED;
1646 uap->sysmsg_result = AIO_CANCELED;
1649 uap->sysmsg_result = AIO_ALLDONE;
1652 #endif /* VFS_AIO */
1656 * aio_error is implemented in the kernel level for compatibility purposes only.
1657 * For a user mode async implementation, it would be best to do it in a userland
1661 sys_aio_error(struct aio_error_args *uap)
1666 struct proc *p = curproc;
1667 struct aiocblist *cb;
1668 struct kaioinfo *ki;
1675 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1676 if ((jobref == -1) || (jobref == 0))
1679 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1680 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1682 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1689 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1691 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1693 uap->sysmsg_result = EINPROGRESS;
1699 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1701 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1703 uap->sysmsg_result = EINPROGRESS;
1711 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1713 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1715 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1721 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1723 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1725 uap->sysmsg_result = EINPROGRESS;
1736 status = fuword(&uap->aiocbp->_aiocb_private.status);
1738 return fuword(&uap->aiocbp->_aiocb_private.error);
1741 #endif /* VFS_AIO */
1744 /* syscall - asynchronous read from a file (REALTIME) */
1746 sys_aio_read(struct aio_read_args *uap)
1751 return aio_aqueue(uap->aiocbp, LIO_READ);
1752 #endif /* VFS_AIO */
1755 /* syscall - asynchronous write to a file (REALTIME) */
1757 sys_aio_write(struct aio_write_args *uap)
1762 return aio_aqueue(uap->aiocbp, LIO_WRITE);
1763 #endif /* VFS_AIO */
1766 /* syscall - XXX undocumented */
1768 sys_lio_listio(struct lio_listio_args *uap)
1773 struct proc *p = curproc;
1774 struct lwp *lp = curthread->td_lwp;
1775 int nent, nentqueued;
1776 struct aiocb *iocb, * const *cbptr;
1777 struct aiocblist *cb;
1778 struct kaioinfo *ki;
1779 struct aio_liojob *lj;
1780 int error, runningcode;
1784 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1788 if (nent > AIO_LISTIO_MAX)
1791 if (p->p_aioinfo == NULL)
1792 aio_init_aioinfo(p);
1794 if ((nent + num_queue_count) > max_queue_count)
1798 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1801 lj = zalloc(aiolio_zone);
1806 lj->lioj_buffer_count = 0;
1807 lj->lioj_buffer_finished_count = 0;
1808 lj->lioj_queue_count = 0;
1809 lj->lioj_queue_finished_count = 0;
1815 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1816 error = copyin(uap->sig, &lj->lioj_signal,
1817 sizeof(lj->lioj_signal));
1819 zfree(aiolio_zone, lj);
1822 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1823 zfree(aiolio_zone, lj);
1826 lj->lioj_flags |= LIOJ_SIGNAL;
1827 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1829 lj->lioj_flags &= ~LIOJ_SIGNAL;
1831 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1833 * Get pointers to the list of I/O requests.
1837 cbptr = uap->acb_list;
1838 for (i = 0; i < uap->nent; i++) {
1839 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1840 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1841 error = _aio_aqueue(iocb, lj, 0);
1850 * If we haven't queued any, then just return error.
1852 if (nentqueued == 0)
1856 * Calculate the appropriate error return.
1862 if (uap->mode == LIO_WAIT) {
1863 int command, found, jobref;
1867 for (i = 0; i < uap->nent; i++) {
1869 * Fetch address of the control buf pointer in
1872 iocb = (struct aiocb *)
1873 (intptr_t)fuword(&cbptr[i]);
1874 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1879 * Fetch the associated command from user space.
1881 command = fuword(&iocb->aio_lio_opcode);
1882 if (command == LIO_NOP) {
1887 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1889 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1890 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1892 if (cb->uaiocb.aio_lio_opcode
1894 lp->lwp_ru.ru_oublock +=
1896 cb->outputcharge = 0;
1897 } else if (cb->uaiocb.aio_lio_opcode
1899 lp->lwp_ru.ru_inblock +=
1901 cb->inputcharge = 0;
1909 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1910 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1920 * If all I/Os have been disposed of, then we can
1923 if (found == nentqueued)
1926 ki->kaio_flags |= KAIO_WAKEUP;
1927 error = tsleep(p, PCATCH, "aiospn", 0);
1931 else if (error == EWOULDBLOCK)
1937 #endif /* VFS_AIO */
1942 * This is a weird hack so that we can post a signal. It is safe to do so from
1943 * a timeout routine, but *not* from an interrupt routine.
1946 process_signal(void *aioj)
1948 struct aiocblist *aiocbe = aioj;
1949 struct aio_liojob *lj = aiocbe->lio;
1950 struct aiocb *cb = &aiocbe->uaiocb;
1952 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1953 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1954 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1955 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1958 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1959 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
1963 * Interrupt handler for physio, performs the necessary process wakeups, and
1967 aio_physwakeup(struct bio *bio)
1969 struct buf *bp = bio->bio_buf;
1970 struct aiocblist *aiocbe;
1972 struct kaioinfo *ki;
1973 struct aio_liojob *lj;
1975 aiocbe = bio->bio_caller_info2.ptr;
1978 p = bio->bio_caller_info1.ptr;
1980 aiocbe->jobstate = JOBST_JOBBFINISHED;
1981 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
1982 aiocbe->uaiocb._aiocb_private.error = 0;
1983 aiocbe->jobflags |= AIOCBLIST_DONE;
1985 if (bp->b_flags & B_ERROR)
1986 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1990 lj->lioj_buffer_finished_count++;
1993 * wakeup/signal if all of the interrupt jobs are done.
1995 if (lj->lioj_buffer_finished_count ==
1996 lj->lioj_buffer_count) {
1998 * Post a signal if it is called for.
2000 if ((lj->lioj_flags &
2001 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2003 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2004 callout_reset(&aiocbe->timeout, 0,
2005 process_signal, aiocbe);
2012 ki->kaio_buffer_finished_count++;
2013 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2014 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2015 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2017 KNOTE(&aiocbe->klist, 0);
2018 /* Do the wakeup. */
2019 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2020 ki->kaio_flags &= ~KAIO_WAKEUP;
2025 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2026 callout_reset(&aiocbe->timeout, 0,
2027 process_signal, aiocbe);
2030 bp->b_cmd = BUF_CMD_DONE;
2033 #endif /* VFS_AIO */
2035 /* syscall - wait for the next completion of an aio request */
2037 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2042 struct proc *p = curproc;
2043 struct lwp *lp = curthread->td_lwp;
2046 struct kaioinfo *ki;
2047 struct aiocblist *cb = NULL;
2050 suword(uap->aiocbp, (int)NULL);
2054 /* Get timespec struct. */
2055 error = copyin(uap->timeout, &ts, sizeof(ts));
2059 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2062 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2063 if (itimerfix(&atv))
2065 timo = tvtohz_high(&atv);
2073 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2074 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2075 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2076 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2077 lp->lwp_ru.ru_oublock +=
2079 cb->outputcharge = 0;
2080 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2081 lp->lwp_ru.ru_inblock += cb->inputcharge;
2082 cb->inputcharge = 0;
2085 return cb->uaiocb._aiocb_private.error;
2089 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2091 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2092 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2094 return cb->uaiocb._aiocb_private.error;
2097 ki->kaio_flags |= KAIO_WAKEUP;
2098 error = tsleep(p, PCATCH, "aiowc", timo);
2101 if (error == ERESTART)
2105 else if (error == EINTR)
2107 else if (error == EWOULDBLOCK)
2110 #endif /* VFS_AIO */
2115 filt_aioattach(struct knote *kn)
2121 struct filterops aio_filtops =
2122 { 0, filt_aioattach, NULL, NULL };
2125 /* kqueue attach function */
2127 filt_aioattach(struct knote *kn)
2129 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2132 * The aiocbe pointer must be validated before using it, so
2133 * registration is restricted to the kernel; the user cannot
2136 if ((kn->kn_flags & EV_FLAG1) == 0)
2138 kn->kn_flags &= ~EV_FLAG1;
2140 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2145 /* kqueue detach function */
2147 filt_aiodetach(struct knote *kn)
2149 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2151 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2154 /* kqueue filter function */
2157 filt_aio(struct knote *kn, long hint)
2159 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2161 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2162 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2163 aiocbe->jobstate != JOBST_JOBBFINISHED)
2165 kn->kn_flags |= EV_EOF;
2169 struct filterops aio_filtops =
2170 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2171 #endif /* VFS_AIO */