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>
51 #include <sys/file2.h>
53 #include <sys/sysref2.h>
54 #include <sys/thread2.h>
55 #include <sys/mplock2.h>
57 #include <machine/limits.h>
58 #include "opt_vfs_aio.h"
63 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
68 #define JOBST_NULL 0x0
69 #define JOBST_JOBQGLOBAL 0x2
70 #define JOBST_JOBRUNNING 0x3
71 #define JOBST_JOBFINISHED 0x4
72 #define JOBST_JOBQBUF 0x5
73 #define JOBST_JOBBFINISHED 0x6
75 #ifndef MAX_AIO_PER_PROC
76 #define MAX_AIO_PER_PROC 32
79 #ifndef MAX_AIO_QUEUE_PER_PROC
80 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
84 #define MAX_AIO_PROCS 32
88 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
91 #ifndef TARGET_AIO_PROCS
92 #define TARGET_AIO_PROCS 4
96 #define MAX_BUF_AIO 16
99 #ifndef AIOD_TIMEOUT_DEFAULT
100 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
103 #ifndef AIOD_LIFETIME_DEFAULT
104 #define AIOD_LIFETIME_DEFAULT (30 * hz)
107 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
109 static int max_aio_procs = MAX_AIO_PROCS;
110 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
111 CTLFLAG_RW, &max_aio_procs, 0,
112 "Maximum number of kernel threads to use for handling async IO");
114 static int num_aio_procs = 0;
115 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
116 CTLFLAG_RD, &num_aio_procs, 0,
117 "Number of presently active kernel threads for async IO");
120 * The code will adjust the actual number of AIO processes towards this
121 * number when it gets a chance.
123 static int target_aio_procs = TARGET_AIO_PROCS;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
125 0, "Preferred number of ready kernel threads for async IO");
127 static int max_queue_count = MAX_AIO_QUEUE;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
129 "Maximum number of aio requests to queue, globally");
131 static int num_queue_count = 0;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
133 "Number of queued aio requests");
135 static int num_buf_aio = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
137 "Number of aio requests presently handled by the buf subsystem");
139 /* Number of async I/O thread in the process of being started */
140 /* XXX This should be local to _aio_aqueue() */
141 static int num_aio_resv_start = 0;
143 static int aiod_timeout;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
145 "Timeout value for synchronous aio operations");
147 static int aiod_lifetime;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
149 "Maximum lifetime for idle aiod");
151 static int max_aio_per_proc = MAX_AIO_PER_PROC;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
153 0, "Maximum active aio requests per process (stored in the process)");
155 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
157 &max_aio_queue_per_proc, 0,
158 "Maximum queued aio requests per process (stored in the process)");
160 static int max_buf_aio = MAX_BUF_AIO;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
162 "Maximum buf aio requests per process (stored in the process)");
167 #define AIOP_FREE 0x1 /* proc on free queue */
168 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
171 int aioprocflags; /* AIO proc flags */
172 TAILQ_ENTRY(aioproclist) list; /* List of processes */
173 struct proc *aioproc; /* The AIO thread */
177 * data-structure for lio signal management
181 int lioj_buffer_count;
182 int lioj_buffer_finished_count;
183 int lioj_queue_count;
184 int lioj_queue_finished_count;
185 struct sigevent lioj_signal; /* signal on all I/O done */
186 TAILQ_ENTRY(aio_liojob) lioj_list;
187 struct kaioinfo *lioj_ki;
189 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
190 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
193 * per process aio data structure
196 int kaio_flags; /* per process kaio flags */
197 int kaio_maxactive_count; /* maximum number of AIOs */
198 int kaio_active_count; /* number of currently used AIOs */
199 int kaio_qallowed_count; /* maxiumu size of AIO queue */
200 int kaio_queue_count; /* size of AIO queue */
201 int kaio_ballowed_count; /* maximum number of buffers */
202 int kaio_queue_finished_count; /* number of daemon jobs finished */
203 int kaio_buffer_count; /* number of physio buffers */
204 int kaio_buffer_finished_count; /* count of I/O done */
205 struct proc *kaio_p; /* process that uses this kaio block */
206 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
207 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
208 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
209 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
210 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
211 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
214 #define KAIO_RUNDOWN 0x1 /* process is being run down */
215 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
217 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
218 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
219 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
220 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
222 static void aio_init_aioinfo(struct proc *p);
223 static void aio_onceonly(void *);
224 static int aio_free_entry(struct aiocblist *aiocbe);
225 static void aio_process(struct aiocblist *aiocbe);
226 static int aio_newproc(void);
227 static int aio_aqueue(struct aiocb *job, int type);
228 static void aio_physwakeup(struct bio *bio);
229 static int aio_fphysio(struct aiocblist *aiocbe);
230 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
231 static void aio_daemon(void *uproc, struct trapframe *frame);
232 static void process_signal(void *aioj);
234 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
238 * kaio Per process async io info
239 * aiop async io thread data
240 * aiocb async io jobs
241 * aiol list io job pointer - internal to aio_suspend XXX
242 * aiolio list io jobs
244 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
247 * Startup initialization
250 aio_onceonly(void *na)
252 TAILQ_INIT(&aio_freeproc);
253 TAILQ_INIT(&aio_activeproc);
254 TAILQ_INIT(&aio_jobs);
255 TAILQ_INIT(&aio_bufjobs);
256 TAILQ_INIT(&aio_freejobs);
257 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
258 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
259 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
260 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
261 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
262 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
263 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
268 * Init the per-process aioinfo structure. The aioinfo limits are set
269 * per-process for user limit (resource) management.
272 aio_init_aioinfo(struct proc *p)
275 if (p->p_aioinfo == NULL) {
276 ki = zalloc(kaio_zone);
279 ki->kaio_maxactive_count = max_aio_per_proc;
280 ki->kaio_active_count = 0;
281 ki->kaio_qallowed_count = max_aio_queue_per_proc;
282 ki->kaio_queue_count = 0;
283 ki->kaio_ballowed_count = max_buf_aio;
284 ki->kaio_buffer_count = 0;
285 ki->kaio_buffer_finished_count = 0;
287 TAILQ_INIT(&ki->kaio_jobdone);
288 TAILQ_INIT(&ki->kaio_jobqueue);
289 TAILQ_INIT(&ki->kaio_bufdone);
290 TAILQ_INIT(&ki->kaio_bufqueue);
291 TAILQ_INIT(&ki->kaio_liojoblist);
292 TAILQ_INIT(&ki->kaio_sockqueue);
295 while (num_aio_procs < target_aio_procs)
300 * Free a job entry. Wait for completion if it is currently active, but don't
301 * delay forever. If we delay, we return a flag that says that we have to
302 * restart the queue scan.
305 aio_free_entry(struct aiocblist *aiocbe)
308 struct aio_liojob *lj;
312 if (aiocbe->jobstate == JOBST_NULL)
313 panic("aio_free_entry: freeing already free job");
315 p = aiocbe->userproc;
319 panic("aio_free_entry: missing p->p_aioinfo");
321 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
322 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
323 tsleep(aiocbe, 0, "jobwai", 0);
325 if (aiocbe->bp == NULL) {
326 if (ki->kaio_queue_count <= 0)
327 panic("aio_free_entry: process queue size <= 0");
328 if (num_queue_count <= 0)
329 panic("aio_free_entry: system wide queue size <= 0");
332 lj->lioj_queue_count--;
333 if (aiocbe->jobflags & AIOCBLIST_DONE)
334 lj->lioj_queue_finished_count--;
336 ki->kaio_queue_count--;
337 if (aiocbe->jobflags & AIOCBLIST_DONE)
338 ki->kaio_queue_finished_count--;
342 lj->lioj_buffer_count--;
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 lj->lioj_buffer_finished_count--;
346 if (aiocbe->jobflags & AIOCBLIST_DONE)
347 ki->kaio_buffer_finished_count--;
348 ki->kaio_buffer_count--;
352 /* aiocbe is going away, we need to destroy any knotes */
353 /* XXX lwp knote wants a thread, but only cares about the process */
354 knote_empty(&aiocbe->klist);
356 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
357 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
358 ki->kaio_flags &= ~KAIO_WAKEUP;
362 if (aiocbe->jobstate == JOBST_JOBQBUF) {
363 if ((error = aio_fphysio(aiocbe)) != 0)
365 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
366 panic("aio_free_entry: invalid physio finish-up state");
368 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
370 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
372 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
373 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
375 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
376 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
377 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
379 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
382 vunmapbuf(aiocbe->bp);
383 relpbuf(aiocbe->bp, NULL);
387 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
388 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
389 zfree(aiolio_zone, lj);
391 aiocbe->jobstate = JOBST_NULL;
392 callout_stop(&aiocbe->timeout);
393 fdrop(aiocbe->fd_file);
394 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
400 * Rundown the jobs for a given process.
403 aio_proc_rundown(struct proc *p)
409 struct aio_liojob *lj, *ljn;
410 struct aiocblist *aiocbe, *aiocbn;
418 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
419 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
420 ki->kaio_buffer_finished_count)) {
421 ki->kaio_flags |= KAIO_RUNDOWN;
422 if (tsleep(p, 0, "kaiowt", aiod_timeout))
427 * Move any aio ops that are waiting on socket I/O to the normal job
428 * queues so they are cleaned up with any others.
431 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
433 aiocbn = TAILQ_NEXT(aiocbe, plist);
434 fp = aiocbe->fd_file;
436 so = (struct socket *)fp->f_data;
437 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
438 if (TAILQ_EMPTY(&so->so_aiojobq)) {
439 so->so_snd.ssb_flags &= ~SSB_AIO;
440 so->so_rcv.ssb_flags &= ~SSB_AIO;
443 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
444 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
445 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
450 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
451 aiocbn = TAILQ_NEXT(aiocbe, plist);
452 if (aio_free_entry(aiocbe))
457 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
459 aiocbn = TAILQ_NEXT(aiocbe, plist);
460 if (aio_free_entry(aiocbe))
466 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
467 ki->kaio_flags |= KAIO_WAKEUP;
468 tsleep(p, 0, "aioprn", 0);
476 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
477 aiocbn = TAILQ_NEXT(aiocbe, plist);
478 if (aio_free_entry(aiocbe)) {
486 * If we've slept, jobs might have moved from one queue to another.
487 * Retry rundown if we didn't manage to empty the queues.
489 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
490 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
491 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
492 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
495 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
496 ljn = TAILQ_NEXT(lj, lioj_list);
497 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
499 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
500 zfree(aiolio_zone, lj);
503 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
504 "QF:%d\n", lj->lioj_buffer_count,
505 lj->lioj_buffer_finished_count,
506 lj->lioj_queue_count,
507 lj->lioj_queue_finished_count);
512 zfree(kaio_zone, ki);
519 * Select a job to run (called by an AIO daemon).
521 static struct aiocblist *
522 aio_selectjob(struct aioproclist *aiop)
524 struct aiocblist *aiocbe;
529 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
530 TAILQ_NEXT(aiocbe, list)) {
531 userp = aiocbe->userproc;
532 ki = userp->p_aioinfo;
534 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
535 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
546 * The AIO processing activity. This is the code that does the I/O request for
547 * the non-physio version of the operations. The normal vn operations are used,
548 * and this code should work in all instances for every type of file, including
549 * pipes, sockets, fifos, and regular files.
552 aio_process(struct aiocblist *aiocbe)
561 int oublock_st, oublock_end;
562 int inblock_st, inblock_end;
565 cb = &aiocbe->uaiocb;
566 fp = aiocbe->fd_file;
568 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
569 aiov.iov_len = cb->aio_nbytes;
571 auio.uio_iov = &aiov;
573 auio.uio_offset = cb->aio_offset;
574 auio.uio_resid = cb->aio_nbytes;
575 cnt = cb->aio_nbytes;
576 auio.uio_segflg = UIO_USERSPACE;
579 inblock_st = mytd->td_lwp->lwp_ru.ru_inblock;
580 oublock_st = mytd->td_lwp->lwp_ru.ru_oublock;
582 * _aio_aqueue() acquires a reference to the file that is
583 * released in aio_free_entry().
585 if (cb->aio_lio_opcode == LIO_READ) {
586 auio.uio_rw = UIO_READ;
587 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET);
589 auio.uio_rw = UIO_WRITE;
590 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET);
592 inblock_end = mytd->td_lwp->lwp_ru.ru_inblock;
593 oublock_end = mytd->td_lwp->lwp_ru.ru_oublock;
595 aiocbe->inputcharge = inblock_end - inblock_st;
596 aiocbe->outputcharge = oublock_end - oublock_st;
598 if ((error) && (auio.uio_resid != cnt)) {
599 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
601 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
602 ksignal(aiocbe->userproc, SIGPIPE);
605 cnt -= auio.uio_resid;
606 cb->_aiocb_private.error = error;
607 cb->_aiocb_private.status = cnt;
611 * The AIO daemon, most of the actual work is done in aio_process,
612 * but the setup (and address space mgmt) is done in this routine.
614 * The MP lock is held on entry.
617 aio_daemon(void *uproc, struct trapframe *frame)
619 struct aio_liojob *lj;
621 struct aiocblist *aiocbe;
622 struct aioproclist *aiop;
624 struct proc *mycp, *userp;
625 struct vmspace *curvm;
629 mylwp = curthread->td_lwp;
630 mycp = mylwp->lwp_proc;
632 if (mycp->p_textvp) {
633 vrele(mycp->p_textvp);
634 mycp->p_textvp = NULL;
638 * Allocate and ready the aio control info. There is one aiop structure
641 aiop = zalloc(aiop_zone);
642 aiop->aioproc = mycp;
643 aiop->aioprocflags |= AIOP_FREE;
648 * Place thread (lightweight process) onto the AIO free thread list.
650 if (TAILQ_EMPTY(&aio_freeproc))
651 wakeup(&aio_freeproc);
652 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
656 /* Make up a name for the daemon. */
657 strcpy(mycp->p_comm, "aiod");
660 * Get rid of our current filedescriptors. AIOD's don't need any
661 * filedescriptors, except as temporarily inherited from the client.
662 * Credentials are also cloned, and made equivalent to "root".
665 cr = cratom(&mycp->p_ucred);
667 uireplace(&cr->cr_uidinfo, uifind(0));
669 cr->cr_groups[0] = 1;
671 /* The daemon resides in its own pgrp. */
672 enterpgrp(mycp, mycp->p_pid, 1);
674 /* Mark special process type. */
675 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
678 * Wakeup parent process. (Parent sleeps to keep from blasting away
679 * and creating too many daemons.)
686 * Take daemon off of free queue
688 if (aiop->aioprocflags & AIOP_FREE) {
690 TAILQ_REMOVE(&aio_freeproc, aiop, list);
691 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
692 aiop->aioprocflags &= ~AIOP_FREE;
695 aiop->aioprocflags &= ~AIOP_SCHED;
700 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
701 cb = &aiocbe->uaiocb;
702 userp = aiocbe->userproc;
704 aiocbe->jobstate = JOBST_JOBRUNNING;
707 * Connect to process address space for user program.
709 if (curvm != userp->p_vmspace) {
710 pmap_setlwpvm(mylwp, userp->p_vmspace);
712 sysref_put(&curvm->vm_sysref);
713 curvm = userp->p_vmspace;
714 sysref_get(&curvm->vm_sysref);
717 ki = userp->p_aioinfo;
720 /* Account for currently active jobs. */
721 ki->kaio_active_count++;
723 /* Do the I/O function. */
726 /* Decrement the active job count. */
727 ki->kaio_active_count--;
730 * Increment the completion count for wakeup/signal
733 aiocbe->jobflags |= AIOCBLIST_DONE;
734 ki->kaio_queue_finished_count++;
736 lj->lioj_queue_finished_count++;
737 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
738 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
739 ki->kaio_flags &= ~KAIO_WAKEUP;
744 if (lj && (lj->lioj_flags &
745 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
746 if ((lj->lioj_queue_finished_count ==
747 lj->lioj_queue_count) &&
748 (lj->lioj_buffer_finished_count ==
749 lj->lioj_buffer_count)) {
751 lj->lioj_signal.sigev_signo);
758 aiocbe->jobstate = JOBST_JOBFINISHED;
761 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
762 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
764 KNOTE(&aiocbe->klist, 0);
766 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
768 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
771 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
772 ksignal(userp, cb->aio_sigevent.sigev_signo);
777 * Disconnect from user address space.
780 /* swap our original address space back in */
781 pmap_setlwpvm(mylwp, mycp->p_vmspace);
782 sysref_put(&curvm->vm_sysref);
787 * If we are the first to be put onto the free queue, wakeup
788 * anyone waiting for a daemon.
791 TAILQ_REMOVE(&aio_activeproc, aiop, list);
792 if (TAILQ_EMPTY(&aio_freeproc))
793 wakeup(&aio_freeproc);
794 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
795 aiop->aioprocflags |= AIOP_FREE;
799 * If daemon is inactive for a long time, allow it to exit,
800 * thereby freeing resources.
802 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
803 0, "aiordy", aiod_lifetime)) {
805 if (TAILQ_EMPTY(&aio_jobs)) {
806 if ((aiop->aioprocflags & AIOP_FREE) &&
807 (num_aio_procs > target_aio_procs)) {
808 TAILQ_REMOVE(&aio_freeproc, aiop, list);
810 zfree(aiop_zone, aiop);
813 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
814 kprintf("AIOD: bad vm refcnt for"
815 " exiting daemon: %d\n",
816 mycp->p_vmspace->vm_sysref.refcnt);
828 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
829 * AIO daemon modifies its environment itself.
835 struct lwp *lp, *nlp;
839 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
842 nlp = ONLY_LWP_IN_PROC(np);
843 cpu_set_fork_handler(nlp, aio_daemon, curproc);
844 start_forked_proc(lp, np);
847 * Wait until daemon is started, but continue on just in case to
848 * handle error conditions.
850 error = tsleep(np, 0, "aiosta", aiod_timeout);
857 * Try the high-performance, low-overhead physio method for eligible
858 * VCHR devices. This method doesn't use an aio helper thread, and
859 * thus has very low overhead.
861 * Assumes that the caller, _aio_aqueue(), has incremented the file
862 * structure's reference count, preventing its deallocation for the
863 * duration of this call.
866 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
874 struct aio_liojob *lj;
877 cb = &aiocbe->uaiocb;
878 fp = aiocbe->fd_file;
880 if (fp->f_type != DTYPE_VNODE)
883 vp = (struct vnode *)fp->f_data;
886 * If its not a disk, we don't want to return a positive error.
887 * It causes the aio code to not fall through to try the thread
888 * way when you're talking to a regular file.
890 if (!vn_isdisk(vp, &error)) {
891 if (error == ENOTBLK)
897 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
901 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
905 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
908 ki->kaio_buffer_count++;
912 lj->lioj_buffer_count++;
914 /* Create and build a buffer header for a transfer. */
919 * Get a copy of the kva from the physical buffer.
921 bp->b_bio1.bio_caller_info1.ptr = p;
924 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
925 BUF_CMD_WRITE : BUF_CMD_READ;
926 bp->b_bio1.bio_done = aio_physwakeup;
927 bp->b_bio1.bio_flags |= BIO_SYNC;
928 bp->b_bio1.bio_offset = cb->aio_offset;
930 /* Bring buffer into kernel space. */
931 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
939 bp->b_bio1.bio_caller_info2.ptr = aiocbe;
940 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
941 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
942 aiocbe->jobstate = JOBST_JOBQBUF;
943 cb->_aiocb_private.status = cb->aio_nbytes;
950 * Perform the transfer. vn_strategy must be used even though we
951 * know we have a device in order to deal with requests which exceed
952 * device DMA limitations.
954 vn_strategy(vp, &bp->b_bio1);
961 * If we had an error invoking the request, or an error in processing
962 * the request before we have returned, we process it as an error in
963 * transfer. Note that such an I/O error is not indicated immediately,
964 * but is returned using the aio_error mechanism. In this case,
965 * aio_suspend will return immediately.
967 if (bp->b_error || (bp->b_flags & B_ERROR)) {
968 struct aiocb *job = aiocbe->uuaiocb;
970 aiocbe->uaiocb._aiocb_private.status = 0;
971 suword(&job->_aiocb_private.status, 0);
972 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
973 suword(&job->_aiocb_private.error, bp->b_error);
975 ki->kaio_buffer_finished_count++;
977 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
978 aiocbe->jobstate = JOBST_JOBBFINISHED;
979 aiocbe->jobflags |= AIOCBLIST_DONE;
980 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
981 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
982 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
989 KNOTE(&aiocbe->klist, 0);
993 ki->kaio_buffer_count--;
995 lj->lioj_buffer_count--;
1002 * This waits/tests physio completion.
1005 aio_fphysio(struct aiocblist *iocb)
1012 error = biowait_timeout(&bp->b_bio1, "physstr", aiod_timeout);
1013 if (error == EWOULDBLOCK)
1016 /* Release mapping into kernel space. */
1022 /* Check for an error. */
1023 if (bp->b_flags & B_ERROR)
1024 error = bp->b_error;
1029 #endif /* VFS_AIO */
1032 * Wake up aio requests that may be serviceable now.
1035 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1040 struct aiocblist *cb,*cbn;
1042 struct kaioinfo *ki = NULL;
1043 int opcode, wakecount = 0;
1044 struct aioproclist *aiop;
1046 if (ssb == &so->so_snd) {
1048 so->so_snd.ssb_flags &= ~SSB_AIO;
1051 so->so_rcv.ssb_flags &= ~SSB_AIO;
1054 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1055 cbn = TAILQ_NEXT(cb, list);
1056 if (opcode == cb->uaiocb.aio_lio_opcode) {
1059 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1060 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1061 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1062 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1064 if (cb->jobstate != JOBST_JOBQGLOBAL)
1065 panic("invalid queue value");
1069 while (wakecount--) {
1070 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1071 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1072 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1073 aiop->aioprocflags &= ~AIOP_FREE;
1074 wakeup(aiop->aioproc);
1077 #endif /* VFS_AIO */
1082 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1083 * technique is done in this code.
1086 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1088 struct proc *p = curproc;
1093 int opcode, user_opcode;
1094 struct aiocblist *aiocbe;
1095 struct aioproclist *aiop;
1096 struct kaioinfo *ki;
1102 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1103 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1105 aiocbe = zalloc (aiocb_zone);
1107 aiocbe->inputcharge = 0;
1108 aiocbe->outputcharge = 0;
1109 callout_init(&aiocbe->timeout);
1110 SLIST_INIT(&aiocbe->klist);
1112 suword(&job->_aiocb_private.status, -1);
1113 suword(&job->_aiocb_private.error, 0);
1114 suword(&job->_aiocb_private.kernelinfo, -1);
1116 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1118 suword(&job->_aiocb_private.error, error);
1119 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1122 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1123 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1124 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1128 /* Save userspace address of the job info. */
1129 aiocbe->uuaiocb = job;
1131 /* Get the opcode. */
1132 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1133 if (type != LIO_NOP)
1134 aiocbe->uaiocb.aio_lio_opcode = type;
1135 opcode = aiocbe->uaiocb.aio_lio_opcode;
1138 * Range check file descriptor.
1140 fflags = (opcode == LIO_WRITE) ? FWRITE : FREAD;
1141 fd = aiocbe->uaiocb.aio_fildes;
1142 fp = holdfp(p->p_fd, fd, fflags);
1144 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1146 suword(&job->_aiocb_private.error, EBADF);
1150 aiocbe->fd_file = fp;
1152 if (aiocbe->uaiocb.aio_offset == -1LL) {
1156 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1161 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1162 if (jobrefid == LONG_MAX)
1167 if (opcode == LIO_NOP) {
1169 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1171 suword(&job->_aiocb_private.error, 0);
1172 suword(&job->_aiocb_private.status, 0);
1173 suword(&job->_aiocb_private.kernelinfo, 0);
1177 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1179 suword(&job->_aiocb_private.status, 0);
1184 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1185 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1186 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1190 * This method for requesting kevent-based notification won't
1191 * work on the alpha, since we're passing in a pointer
1192 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1193 * based method instead.
1195 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1196 user_opcode == LIO_WRITE)
1199 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1204 kq_fp = holdfp(p->p_fd, (int)kev.ident, -1);
1205 if (kq_fp == NULL || kq_fp->f_type != DTYPE_KQUEUE) {
1213 kq = (struct kqueue *)kq_fp->f_data;
1214 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1215 kev.filter = EVFILT_AIO;
1216 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1217 kev.data = (intptr_t)aiocbe;
1218 error = kqueue_register(kq, &kev);
1223 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1225 suword(&job->_aiocb_private.error, error);
1230 suword(&job->_aiocb_private.error, EINPROGRESS);
1231 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1232 aiocbe->userproc = p;
1233 aiocbe->jobflags = 0;
1237 if (fp->f_type == DTYPE_SOCKET) {
1239 * Alternate queueing for socket ops: Reach down into the
1240 * descriptor to get the socket data. Then check to see if the
1241 * socket is ready to be read or written (based on the requested
1244 * If it is not ready for io, then queue the aiocbe on the
1245 * socket, and set the flags so we get a call when ssb_notify()
1248 so = (struct socket *)fp->f_data;
1250 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1251 LIO_WRITE) && (!sowriteable(so)))) {
1252 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1253 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1254 if (opcode == LIO_READ)
1255 so->so_rcv.ssb_flags |= SSB_AIO;
1257 so->so_snd.ssb_flags |= SSB_AIO;
1258 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1259 ki->kaio_queue_count++;
1268 if ((error = aio_qphysio(p, aiocbe)) == 0)
1271 suword(&job->_aiocb_private.status, 0);
1272 aiocbe->uaiocb._aiocb_private.error = error;
1273 suword(&job->_aiocb_private.error, error);
1277 /* No buffer for daemon I/O. */
1280 ki->kaio_queue_count++;
1282 lj->lioj_queue_count++;
1284 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1285 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1287 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1293 * If we don't have a free AIO process, and we are below our quota, then
1294 * start one. Otherwise, depend on the subsequent I/O completions to
1295 * pick-up this job. If we don't successfully create the new process
1296 * (thread) due to resource issues, we return an error for now (EAGAIN),
1297 * which is likely not the correct thing to do.
1301 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1302 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1303 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1304 aiop->aioprocflags &= ~AIOP_FREE;
1305 wakeup(aiop->aioproc);
1306 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1307 ((ki->kaio_active_count + num_aio_resv_start) <
1308 ki->kaio_maxactive_count)) {
1309 num_aio_resv_start++;
1310 if ((error = aio_newproc()) == 0) {
1311 num_aio_resv_start--;
1314 num_aio_resv_start--;
1322 * This routine queues an AIO request, checking for quotas.
1325 aio_aqueue(struct aiocb *job, int type)
1327 struct proc *p = curproc;
1328 struct kaioinfo *ki;
1330 if (p->p_aioinfo == NULL)
1331 aio_init_aioinfo(p);
1333 if (num_queue_count >= max_queue_count)
1337 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1340 return _aio_aqueue(job, NULL, type);
1342 #endif /* VFS_AIO */
1345 * Support the aio_return system call, as a side-effect, kernel resources are
1351 sys_aio_return(struct aio_return_args *uap)
1356 struct proc *p = curproc;
1357 struct lwp *lp = curthread->td_lwp;
1359 struct aiocblist *cb, *ncb;
1361 struct kaioinfo *ki;
1370 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1371 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;
1384 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1385 lp->lwp_ru.ru_oublock += cb->outputcharge;
1386 cb->outputcharge = 0;
1387 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1388 lp->lwp_ru.ru_inblock += cb->inputcharge;
1389 cb->inputcharge = 0;
1397 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1398 ncb = TAILQ_NEXT(cb, plist);
1399 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1402 if (ujob == cb->uuaiocb) {
1403 uap->sysmsg_result =
1404 cb->uaiocb._aiocb_private.status;
1406 uap->sysmsg_result = EFAULT;
1418 #endif /* VFS_AIO */
1422 * Allow a process to wakeup when any of the I/O requests are completed.
1427 sys_aio_suspend(struct aio_suspend_args *uap)
1432 struct proc *p = curproc;
1435 struct aiocb *const *cbptr, *cbp;
1436 struct kaioinfo *ki;
1437 struct aiocblist *cb;
1442 struct aiocb **ujoblist;
1444 if ((u_int)uap->nent > AIO_LISTIO_MAX)
1449 /* Get timespec struct. */
1450 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1453 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1456 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1457 if (itimerfix(&atv))
1459 timo = tvtohz_high(&atv);
1469 ijoblist = zalloc(aiol_zone);
1470 ujoblist = zalloc(aiol_zone);
1471 cbptr = uap->aiocbp;
1473 for (i = 0; i < uap->nent; i++) {
1474 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1477 ujoblist[njoblist] = cbp;
1478 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1482 if (njoblist == 0) {
1483 zfree(aiol_zone, ijoblist);
1484 zfree(aiol_zone, ujoblist);
1491 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1492 for (i = 0; i < njoblist; i++) {
1494 cb->uaiocb._aiocb_private.kernelinfo) ==
1496 if (ujoblist[i] != cb->uuaiocb)
1498 zfree(aiol_zone, ijoblist);
1499 zfree(aiol_zone, ujoblist);
1506 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1507 TAILQ_NEXT(cb, plist)) {
1508 for (i = 0; i < njoblist; i++) {
1510 cb->uaiocb._aiocb_private.kernelinfo) ==
1513 if (ujoblist[i] != cb->uuaiocb)
1515 zfree(aiol_zone, ijoblist);
1516 zfree(aiol_zone, ujoblist);
1522 ki->kaio_flags |= KAIO_WAKEUP;
1523 error = tsleep(p, PCATCH, "aiospn", timo);
1526 if (error == ERESTART || error == EINTR) {
1527 zfree(aiol_zone, ijoblist);
1528 zfree(aiol_zone, ujoblist);
1531 } else if (error == EWOULDBLOCK) {
1532 zfree(aiol_zone, ijoblist);
1533 zfree(aiol_zone, ujoblist);
1544 #endif /* VFS_AIO */
1548 * aio_cancel cancels any non-physio aio operations not currently in
1554 sys_aio_cancel(struct aio_cancel_args *uap)
1559 struct proc *p = curproc;
1560 struct kaioinfo *ki;
1561 struct aiocblist *cbe, *cbn;
1570 fp = holdfp(p->p_fd, uap->fd, -1);
1576 if (fp->f_type == DTYPE_VNODE) {
1577 vp = (struct vnode *)fp->f_data;
1579 if (vn_isdisk(vp,&error)) {
1580 uap->sysmsg_result = AIO_NOTCANCELED;
1584 } else if (fp->f_type == DTYPE_SOCKET) {
1585 so = (struct socket *)fp->f_data;
1589 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1590 cbn = TAILQ_NEXT(cbe, list);
1591 if ((uap->aiocbp == NULL) ||
1592 (uap->aiocbp == cbe->uuaiocb) ) {
1595 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1596 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1597 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1598 if (ki->kaio_flags & KAIO_WAKEUP) {
1601 cbe->jobstate = JOBST_JOBFINISHED;
1602 cbe->uaiocb._aiocb_private.status=-1;
1603 cbe->uaiocb._aiocb_private.error=ECANCELED;
1605 /* XXX cancelled, knote? */
1606 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1608 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1615 if ((cancelled) && (uap->aiocbp)) {
1616 uap->sysmsg_result = AIO_CANCELED;
1626 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1627 cbn = TAILQ_NEXT(cbe, plist);
1629 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1630 ((uap->aiocbp == NULL ) ||
1631 (uap->aiocbp == cbe->uuaiocb))) {
1633 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1634 TAILQ_REMOVE(&aio_jobs, cbe, list);
1635 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1636 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1639 ki->kaio_queue_finished_count++;
1640 cbe->jobstate = JOBST_JOBFINISHED;
1641 cbe->uaiocb._aiocb_private.status = -1;
1642 cbe->uaiocb._aiocb_private.error = ECANCELED;
1643 /* XXX cancelled, knote? */
1644 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1646 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1655 uap->sysmsg_result = AIO_NOTCANCELED;
1657 uap->sysmsg_result = AIO_CANCELED;
1659 uap->sysmsg_result = AIO_ALLDONE;
1665 #endif /* VFS_AIO */
1669 * aio_error is implemented in the kernel level for compatibility purposes only.
1670 * For a user mode async implementation, it would be best to do it in a userland
1676 sys_aio_error(struct aio_error_args *uap)
1681 struct proc *p = curproc;
1682 struct aiocblist *cb;
1683 struct kaioinfo *ki;
1691 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1692 if ((jobref == -1) || (jobref == 0))
1698 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1699 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1701 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1708 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1710 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1712 uap->sysmsg_result = EINPROGRESS;
1718 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1720 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1722 uap->sysmsg_result = EINPROGRESS;
1730 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1732 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1734 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1740 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1742 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1744 uap->sysmsg_result = EINPROGRESS;
1754 #endif /* VFS_AIO */
1758 * syscall - asynchronous read from a file (REALTIME)
1763 sys_aio_read(struct aio_read_args *uap)
1771 error = aio_aqueue(uap->aiocbp, LIO_READ);
1774 #endif /* VFS_AIO */
1778 * syscall - asynchronous write to a file (REALTIME)
1783 sys_aio_write(struct aio_write_args *uap)
1791 error = aio_aqueue(uap->aiocbp, LIO_WRITE);
1794 #endif /* VFS_AIO */
1798 * syscall - XXX undocumented
1803 sys_lio_listio(struct lio_listio_args *uap)
1808 struct proc *p = curproc;
1809 struct lwp *lp = curthread->td_lwp;
1810 int nent, nentqueued;
1811 struct aiocb *iocb, * const *cbptr;
1812 struct aiocblist *cb;
1813 struct kaioinfo *ki;
1814 struct aio_liojob *lj;
1815 int error, runningcode;
1819 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1823 if (nent > AIO_LISTIO_MAX)
1828 if (p->p_aioinfo == NULL)
1829 aio_init_aioinfo(p);
1831 if ((nent + num_queue_count) > max_queue_count) {
1837 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) {
1842 lj = zalloc(aiolio_zone);
1849 lj->lioj_buffer_count = 0;
1850 lj->lioj_buffer_finished_count = 0;
1851 lj->lioj_queue_count = 0;
1852 lj->lioj_queue_finished_count = 0;
1858 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1859 error = copyin(uap->sig, &lj->lioj_signal,
1860 sizeof(lj->lioj_signal));
1862 zfree(aiolio_zone, lj);
1865 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1866 zfree(aiolio_zone, lj);
1870 lj->lioj_flags |= LIOJ_SIGNAL;
1871 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1873 lj->lioj_flags &= ~LIOJ_SIGNAL;
1875 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1877 * Get pointers to the list of I/O requests.
1881 cbptr = uap->acb_list;
1882 for (i = 0; i < uap->nent; i++) {
1883 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1884 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1885 error = _aio_aqueue(iocb, lj, 0);
1894 * If we haven't queued any, then just return error.
1896 if (nentqueued == 0) {
1902 * Calculate the appropriate error return.
1908 if (uap->mode == LIO_WAIT) {
1909 int command, found, jobref;
1913 for (i = 0; i < uap->nent; i++) {
1915 * Fetch address of the control buf pointer in
1918 iocb = (struct aiocb *)
1919 (intptr_t)fuword(&cbptr[i]);
1920 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1925 * Fetch the associated command from user space.
1927 command = fuword(&iocb->aio_lio_opcode);
1928 if (command == LIO_NOP) {
1933 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1935 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1936 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1938 if (cb->uaiocb.aio_lio_opcode
1940 lp->lwp_ru.ru_oublock +=
1942 cb->outputcharge = 0;
1943 } else if (cb->uaiocb.aio_lio_opcode
1945 lp->lwp_ru.ru_inblock +=
1947 cb->inputcharge = 0;
1955 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1956 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1966 * If all I/Os have been disposed of, then we can
1969 if (found == nentqueued) {
1970 error = runningcode;
1974 ki->kaio_flags |= KAIO_WAKEUP;
1975 error = tsleep(p, PCATCH, "aiospn", 0);
1977 if (error == EINTR) {
1979 } else if (error == EWOULDBLOCK) {
1986 error = runningcode;
1990 #endif /* VFS_AIO */
1995 * This is a weird hack so that we can post a signal. It is safe to do so from
1996 * a timeout routine, but *not* from an interrupt routine.
1999 process_signal(void *aioj)
2001 struct aiocblist *aiocbe = aioj;
2002 struct aio_liojob *lj = aiocbe->lio;
2003 struct aiocb *cb = &aiocbe->uaiocb;
2005 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2006 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2007 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2008 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2011 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2012 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2016 * Interrupt handler for physio, performs the necessary process wakeups, and
2020 aio_physwakeup(struct bio *bio)
2022 struct buf *bp = bio->bio_buf;
2023 struct aiocblist *aiocbe;
2025 struct kaioinfo *ki;
2026 struct aio_liojob *lj;
2028 aiocbe = bio->bio_caller_info2.ptr;
2031 p = bio->bio_caller_info1.ptr;
2033 aiocbe->jobstate = JOBST_JOBBFINISHED;
2034 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2035 aiocbe->uaiocb._aiocb_private.error = 0;
2036 aiocbe->jobflags |= AIOCBLIST_DONE;
2038 if (bp->b_flags & B_ERROR)
2039 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2043 lj->lioj_buffer_finished_count++;
2046 * wakeup/signal if all of the interrupt jobs are done.
2048 if (lj->lioj_buffer_finished_count ==
2049 lj->lioj_buffer_count) {
2051 * Post a signal if it is called for.
2053 if ((lj->lioj_flags &
2054 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2056 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2057 callout_reset(&aiocbe->timeout, 0,
2058 process_signal, aiocbe);
2065 ki->kaio_buffer_finished_count++;
2066 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2067 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2068 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2070 KNOTE(&aiocbe->klist, 0);
2071 /* Do the wakeup. */
2072 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2073 ki->kaio_flags &= ~KAIO_WAKEUP;
2078 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2079 callout_reset(&aiocbe->timeout, 0,
2080 process_signal, aiocbe);
2085 #endif /* VFS_AIO */
2088 * syscall - wait for the next completion of an aio request
2093 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2098 struct proc *p = curproc;
2099 struct lwp *lp = curthread->td_lwp;
2102 struct kaioinfo *ki;
2103 struct aiocblist *cb = NULL;
2106 suword(uap->aiocbp, (int)NULL);
2110 /* Get timespec struct. */
2111 error = copyin(uap->timeout, &ts, sizeof(ts));
2115 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2118 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2119 if (itimerfix(&atv))
2121 timo = tvtohz_high(&atv);
2131 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2132 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2133 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2134 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2135 lp->lwp_ru.ru_oublock +=
2137 cb->outputcharge = 0;
2138 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2139 lp->lwp_ru.ru_inblock += cb->inputcharge;
2140 cb->inputcharge = 0;
2143 error = cb->uaiocb._aiocb_private.error;
2148 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2150 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2151 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2153 error = cb->uaiocb._aiocb_private.error;
2157 ki->kaio_flags |= KAIO_WAKEUP;
2158 error = tsleep(p, PCATCH, "aiowc", timo);
2161 if (error == ERESTART) {
2169 if (error == EWOULDBLOCK) {
2176 #endif /* VFS_AIO */
2181 filt_aioattach(struct knote *kn)
2187 struct filterops aio_filtops =
2188 { 0, filt_aioattach, NULL, NULL };
2191 /* kqueue attach function */
2193 filt_aioattach(struct knote *kn)
2195 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2198 * The aiocbe pointer must be validated before using it, so
2199 * registration is restricted to the kernel; the user cannot
2202 if ((kn->kn_flags & EV_FLAG1) == 0)
2204 kn->kn_flags &= ~EV_FLAG1;
2206 knote_insert(&aiocbe->klist, kn);
2211 /* kqueue detach function */
2213 filt_aiodetach(struct knote *kn)
2215 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2217 knote_remove(&aiocbe->klist, kn);
2220 /* kqueue filter function */
2223 filt_aio(struct knote *kn, long hint)
2225 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2227 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2228 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2229 aiocbe->jobstate != JOBST_JOBBFINISHED)
2231 kn->kn_flags |= EV_EOF;
2235 struct filterops aio_filtops =
2236 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2237 #endif /* VFS_AIO */