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 1.6 2003/07/06 21:23:51 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 <machine/limits.h>
52 #include "opt_vfs_aio.h"
57 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
62 #define JOBST_NULL 0x0
63 #define JOBST_JOBQGLOBAL 0x2
64 #define JOBST_JOBRUNNING 0x3
65 #define JOBST_JOBFINISHED 0x4
66 #define JOBST_JOBQBUF 0x5
67 #define JOBST_JOBBFINISHED 0x6
69 #ifndef MAX_AIO_PER_PROC
70 #define MAX_AIO_PER_PROC 32
73 #ifndef MAX_AIO_QUEUE_PER_PROC
74 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
78 #define MAX_AIO_PROCS 32
82 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
85 #ifndef TARGET_AIO_PROCS
86 #define TARGET_AIO_PROCS 4
90 #define MAX_BUF_AIO 16
93 #ifndef AIOD_TIMEOUT_DEFAULT
94 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
97 #ifndef AIOD_LIFETIME_DEFAULT
98 #define AIOD_LIFETIME_DEFAULT (30 * hz)
101 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
103 static int max_aio_procs = MAX_AIO_PROCS;
104 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
105 CTLFLAG_RW, &max_aio_procs, 0,
106 "Maximum number of kernel threads to use for handling async IO");
108 static int num_aio_procs = 0;
109 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
110 CTLFLAG_RD, &num_aio_procs, 0,
111 "Number of presently active kernel threads for async IO");
114 * The code will adjust the actual number of AIO processes towards this
115 * number when it gets a chance.
117 static int target_aio_procs = TARGET_AIO_PROCS;
118 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
119 0, "Preferred number of ready kernel threads for async IO");
121 static int max_queue_count = MAX_AIO_QUEUE;
122 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
123 "Maximum number of aio requests to queue, globally");
125 static int num_queue_count = 0;
126 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
127 "Number of queued aio requests");
129 static int num_buf_aio = 0;
130 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
131 "Number of aio requests presently handled by the buf subsystem");
133 /* Number of async I/O thread in the process of being started */
134 /* XXX This should be local to _aio_aqueue() */
135 static int num_aio_resv_start = 0;
137 static int aiod_timeout;
138 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
139 "Timeout value for synchronous aio operations");
141 static int aiod_lifetime;
142 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
143 "Maximum lifetime for idle aiod");
145 static int max_aio_per_proc = MAX_AIO_PER_PROC;
146 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
147 0, "Maximum active aio requests per process (stored in the process)");
149 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
150 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
151 &max_aio_queue_per_proc, 0,
152 "Maximum queued aio requests per process (stored in the process)");
154 static int max_buf_aio = MAX_BUF_AIO;
155 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
156 "Maximum buf aio requests per process (stored in the process)");
161 #define AIOP_FREE 0x1 /* proc on free queue */
162 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
165 int aioprocflags; /* AIO proc flags */
166 TAILQ_ENTRY(aioproclist) list; /* List of processes */
167 struct proc *aioproc; /* The AIO thread */
171 * data-structure for lio signal management
175 int lioj_buffer_count;
176 int lioj_buffer_finished_count;
177 int lioj_queue_count;
178 int lioj_queue_finished_count;
179 struct sigevent lioj_signal; /* signal on all I/O done */
180 TAILQ_ENTRY(aio_liojob) lioj_list;
181 struct kaioinfo *lioj_ki;
183 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
184 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
187 * per process aio data structure
190 int kaio_flags; /* per process kaio flags */
191 int kaio_maxactive_count; /* maximum number of AIOs */
192 int kaio_active_count; /* number of currently used AIOs */
193 int kaio_qallowed_count; /* maxiumu size of AIO queue */
194 int kaio_queue_count; /* size of AIO queue */
195 int kaio_ballowed_count; /* maximum number of buffers */
196 int kaio_queue_finished_count; /* number of daemon jobs finished */
197 int kaio_buffer_count; /* number of physio buffers */
198 int kaio_buffer_finished_count; /* count of I/O done */
199 struct proc *kaio_p; /* process that uses this kaio block */
200 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
201 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
202 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
203 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
204 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
205 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
208 #define KAIO_RUNDOWN 0x1 /* process is being run down */
209 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
211 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
212 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
213 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
214 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
216 static void aio_init_aioinfo(struct proc *p);
217 static void aio_onceonly(void *);
218 static int aio_free_entry(struct aiocblist *aiocbe);
219 static void aio_process(struct aiocblist *aiocbe);
220 static int aio_newproc(void);
221 static int aio_aqueue(struct aiocb *job, int type);
222 static void aio_physwakeup(struct buf *bp);
223 static int aio_fphysio(struct aiocblist *aiocbe);
224 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
225 static void aio_daemon(void *uproc);
226 static void process_signal(void *aioj);
228 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
232 * kaio Per process async io info
233 * aiop async io thread data
234 * aiocb async io jobs
235 * aiol list io job pointer - internal to aio_suspend XXX
236 * aiolio list io jobs
238 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
241 * Startup initialization
244 aio_onceonly(void *na)
246 TAILQ_INIT(&aio_freeproc);
247 TAILQ_INIT(&aio_activeproc);
248 TAILQ_INIT(&aio_jobs);
249 TAILQ_INIT(&aio_bufjobs);
250 TAILQ_INIT(&aio_freejobs);
251 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
252 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
253 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
254 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
255 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
256 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
257 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
262 * Init the per-process aioinfo structure. The aioinfo limits are set
263 * per-process for user limit (resource) management.
266 aio_init_aioinfo(struct proc *p)
269 if (p->p_aioinfo == NULL) {
270 ki = zalloc(kaio_zone);
273 ki->kaio_maxactive_count = max_aio_per_proc;
274 ki->kaio_active_count = 0;
275 ki->kaio_qallowed_count = max_aio_queue_per_proc;
276 ki->kaio_queue_count = 0;
277 ki->kaio_ballowed_count = max_buf_aio;
278 ki->kaio_buffer_count = 0;
279 ki->kaio_buffer_finished_count = 0;
281 TAILQ_INIT(&ki->kaio_jobdone);
282 TAILQ_INIT(&ki->kaio_jobqueue);
283 TAILQ_INIT(&ki->kaio_bufdone);
284 TAILQ_INIT(&ki->kaio_bufqueue);
285 TAILQ_INIT(&ki->kaio_liojoblist);
286 TAILQ_INIT(&ki->kaio_sockqueue);
289 while (num_aio_procs < target_aio_procs)
294 * Free a job entry. Wait for completion if it is currently active, but don't
295 * delay forever. If we delay, we return a flag that says that we have to
296 * restart the queue scan.
299 aio_free_entry(struct aiocblist *aiocbe)
302 struct aio_liojob *lj;
307 if (aiocbe->jobstate == JOBST_NULL)
308 panic("aio_free_entry: freeing already free job");
310 p = aiocbe->userproc;
314 panic("aio_free_entry: missing p->p_aioinfo");
316 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
317 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
318 tsleep(aiocbe, PRIBIO, "jobwai", 0);
320 if (aiocbe->bp == NULL) {
321 if (ki->kaio_queue_count <= 0)
322 panic("aio_free_entry: process queue size <= 0");
323 if (num_queue_count <= 0)
324 panic("aio_free_entry: system wide queue size <= 0");
327 lj->lioj_queue_count--;
328 if (aiocbe->jobflags & AIOCBLIST_DONE)
329 lj->lioj_queue_finished_count--;
331 ki->kaio_queue_count--;
332 if (aiocbe->jobflags & AIOCBLIST_DONE)
333 ki->kaio_queue_finished_count--;
337 lj->lioj_buffer_count--;
338 if (aiocbe->jobflags & AIOCBLIST_DONE)
339 lj->lioj_buffer_finished_count--;
341 if (aiocbe->jobflags & AIOCBLIST_DONE)
342 ki->kaio_buffer_finished_count--;
343 ki->kaio_buffer_count--;
347 /* aiocbe is going away, we need to destroy any knotes */
348 knote_remove(p, &aiocbe->klist);
350 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
351 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
352 ki->kaio_flags &= ~KAIO_WAKEUP;
356 if (aiocbe->jobstate == JOBST_JOBQBUF) {
357 if ((error = aio_fphysio(aiocbe)) != 0)
359 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
360 panic("aio_free_entry: invalid physio finish-up state");
362 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
364 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
366 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
367 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
369 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
370 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
371 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
373 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
376 vunmapbuf(aiocbe->bp);
377 relpbuf(aiocbe->bp, NULL);
381 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
382 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
383 zfree(aiolio_zone, lj);
385 aiocbe->jobstate = JOBST_NULL;
386 untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
387 fdrop(aiocbe->fd_file, curproc);
388 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
394 * Rundown the jobs for a given process.
397 aio_proc_rundown(struct proc *p)
404 struct aio_liojob *lj, *ljn;
405 struct aiocblist *aiocbe, *aiocbn;
413 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
414 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
415 ki->kaio_buffer_finished_count)) {
416 ki->kaio_flags |= KAIO_RUNDOWN;
417 if (tsleep(p, PRIBIO, "kaiowt", aiod_timeout))
422 * Move any aio ops that are waiting on socket I/O to the normal job
423 * queues so they are cleaned up with any others.
426 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
428 aiocbn = TAILQ_NEXT(aiocbe, plist);
429 fp = aiocbe->fd_file;
431 so = (struct socket *)fp->f_data;
432 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
433 if (TAILQ_EMPTY(&so->so_aiojobq)) {
434 so->so_snd.sb_flags &= ~SB_AIO;
435 so->so_rcv.sb_flags &= ~SB_AIO;
438 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
439 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
440 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
445 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
446 aiocbn = TAILQ_NEXT(aiocbe, plist);
447 if (aio_free_entry(aiocbe))
452 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
454 aiocbn = TAILQ_NEXT(aiocbe, plist);
455 if (aio_free_entry(aiocbe))
460 * Note the use of lots of splbio here, trying to avoid splbio for long chains
461 * of I/O. Probably unnecessary.
465 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
466 ki->kaio_flags |= KAIO_WAKEUP;
467 tsleep(p, PRIBIO, "aioprn", 0);
475 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
476 aiocbn = TAILQ_NEXT(aiocbe, plist);
477 if (aio_free_entry(aiocbe)) {
485 * If we've slept, jobs might have moved from one queue to another.
486 * Retry rundown if we didn't manage to empty the queues.
488 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
489 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
490 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
491 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
494 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
495 ljn = TAILQ_NEXT(lj, lioj_list);
496 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
498 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
499 zfree(aiolio_zone, lj);
502 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
503 "QF:%d\n", lj->lioj_buffer_count,
504 lj->lioj_buffer_finished_count,
505 lj->lioj_queue_count,
506 lj->lioj_queue_finished_count);
511 zfree(kaio_zone, ki);
518 * Select a job to run (called by an AIO daemon).
520 static struct aiocblist *
521 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;
577 auio.uio_procp = mycp;
579 inblock_st = mycp->p_stats->p_ru.ru_inblock;
580 oublock_st = mycp->p_stats->p_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, FOF_OFFSET, mycp);
589 auio.uio_rw = UIO_WRITE;
590 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mycp);
592 inblock_end = mycp->p_stats->p_ru.ru_inblock;
593 oublock_end = mycp->p_stats->p_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 psignal(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)
620 struct aio_liojob *lj;
622 struct aiocblist *aiocbe;
623 struct aioproclist *aiop;
625 struct proc *curcp, *mycp, *userp;
626 struct vmspace *myvm, *tmpvm;
630 * Local copies of curproc (cp) and vmspace (myvm)
633 myvm = mycp->p_vmspace;
635 if (mycp->p_textvp) {
636 vrele(mycp->p_textvp);
637 mycp->p_textvp = NULL;
641 * Allocate and ready the aio control info. There is one aiop structure
644 aiop = zalloc(aiop_zone);
645 aiop->aioproc = mycp;
646 aiop->aioprocflags |= AIOP_FREE;
651 * Place thread (lightweight process) onto the AIO free thread list.
653 if (TAILQ_EMPTY(&aio_freeproc))
654 wakeup(&aio_freeproc);
655 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
659 /* Make up a name for the daemon. */
660 strcpy(mycp->p_comm, "aiod");
663 * Get rid of our current filedescriptors. AIOD's don't need any
664 * filedescriptors, except as temporarily inherited from the client.
665 * Credentials are also cloned, and made equivalent to "root".
669 cr = cratom(&mycp->p_ucred);
671 uifree(cr->cr_uidinfo);
672 cr->cr_uidinfo = uifind(0);
674 cr->cr_groups[0] = 1;
676 /* The daemon resides in its own pgrp. */
677 enterpgrp(mycp, mycp->p_pid, 1);
679 /* Mark special process type. */
680 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
683 * Wakeup parent process. (Parent sleeps to keep from blasting away
684 * and creating too many daemons.)
690 * curcp is the current daemon process context.
691 * userp is the current user process context.
696 * Take daemon off of free queue
698 if (aiop->aioprocflags & AIOP_FREE) {
700 TAILQ_REMOVE(&aio_freeproc, aiop, list);
701 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
702 aiop->aioprocflags &= ~AIOP_FREE;
705 aiop->aioprocflags &= ~AIOP_SCHED;
710 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
711 cb = &aiocbe->uaiocb;
712 userp = aiocbe->userproc;
714 aiocbe->jobstate = JOBST_JOBRUNNING;
717 * Connect to process address space for user program.
719 if (userp != curcp) {
721 * Save the current address space that we are
724 tmpvm = mycp->p_vmspace;
727 * Point to the new user address space, and
730 mycp->p_vmspace = userp->p_vmspace;
731 mycp->p_vmspace->vm_refcnt++;
733 /* Activate the new mapping. */
737 * If the old address space wasn't the daemons
738 * own address space, then we need to remove the
739 * daemon's reference from the other process
740 * that it was acting on behalf of.
748 ki = userp->p_aioinfo;
751 /* Account for currently active jobs. */
752 ki->kaio_active_count++;
754 /* Do the I/O function. */
757 /* Decrement the active job count. */
758 ki->kaio_active_count--;
761 * Increment the completion count for wakeup/signal
764 aiocbe->jobflags |= AIOCBLIST_DONE;
765 ki->kaio_queue_finished_count++;
767 lj->lioj_queue_finished_count++;
768 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
769 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
770 ki->kaio_flags &= ~KAIO_WAKEUP;
775 if (lj && (lj->lioj_flags &
776 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
777 if ((lj->lioj_queue_finished_count ==
778 lj->lioj_queue_count) &&
779 (lj->lioj_buffer_finished_count ==
780 lj->lioj_buffer_count)) {
782 lj->lioj_signal.sigev_signo);
789 aiocbe->jobstate = JOBST_JOBFINISHED;
792 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
793 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
795 KNOTE(&aiocbe->klist, 0);
797 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
799 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
802 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
803 psignal(userp, cb->aio_sigevent.sigev_signo);
808 * Disconnect from user address space.
811 /* Get the user address space to disconnect from. */
812 tmpvm = mycp->p_vmspace;
814 /* Get original address space for daemon. */
815 mycp->p_vmspace = myvm;
817 /* Activate the daemon's address space. */
821 printf("AIOD: vmspace problem -- %d\n",
825 /* Remove our vmspace reference. */
832 * If we are the first to be put onto the free queue, wakeup
833 * anyone waiting for a daemon.
836 TAILQ_REMOVE(&aio_activeproc, aiop, list);
837 if (TAILQ_EMPTY(&aio_freeproc))
838 wakeup(&aio_freeproc);
839 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
840 aiop->aioprocflags |= AIOP_FREE;
844 * If daemon is inactive for a long time, allow it to exit,
845 * thereby freeing resources.
847 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
848 PRIBIO, "aiordy", aiod_lifetime)) {
850 if (TAILQ_EMPTY(&aio_jobs)) {
851 if ((aiop->aioprocflags & AIOP_FREE) &&
852 (num_aio_procs > target_aio_procs)) {
853 TAILQ_REMOVE(&aio_freeproc, aiop, list);
855 zfree(aiop_zone, aiop);
858 if (mycp->p_vmspace->vm_refcnt <= 1) {
859 printf("AIOD: bad vm refcnt for"
860 " exiting daemon: %d\n",
861 mycp->p_vmspace->vm_refcnt);
873 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
874 * AIO daemon modifies its environment itself.
883 error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
886 cpu_set_fork_handler(np, aio_daemon, curproc);
887 start_forked_proc(p, np);
890 * Wait until daemon is started, but continue on just in case to
891 * handle error conditions.
893 error = tsleep(np, PZERO, "aiosta", aiod_timeout);
900 * Try the high-performance, low-overhead physio method for eligible
901 * VCHR devices. This method doesn't use an aio helper thread, and
902 * thus has very low overhead.
904 * Assumes that the caller, _aio_aqueue(), has incremented the file
905 * structure's reference count, preventing its deallocation for the
906 * duration of this call.
909 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
917 struct aio_liojob *lj;
921 cb = &aiocbe->uaiocb;
922 fp = aiocbe->fd_file;
924 if (fp->f_type != DTYPE_VNODE)
927 vp = (struct vnode *)fp->f_data;
930 * If its not a disk, we don't want to return a positive error.
931 * It causes the aio code to not fall through to try the thread
932 * way when you're talking to a regular file.
934 if (!vn_isdisk(vp, &error)) {
935 if (error == ENOTBLK)
941 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
945 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
949 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
952 ki->kaio_buffer_count++;
956 lj->lioj_buffer_count++;
958 /* Create and build a buffer header for a transfer. */
959 bp = (struct buf *)getpbuf(NULL);
963 * Get a copy of the kva from the physical buffer.
966 bp->b_dev = vp->v_rdev;
969 bp->b_bcount = cb->aio_nbytes;
970 bp->b_bufsize = cb->aio_nbytes;
971 bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
973 bp->b_iodone = aio_physwakeup;
974 bp->b_saveaddr = bp->b_data;
975 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
976 bp->b_blkno = btodb(cb->aio_offset);
978 /* Bring buffer into kernel space. */
979 if (vmapbuf(bp) < 0) {
986 bp->b_spc = (void *)aiocbe;
987 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
988 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
989 aiocbe->jobstate = JOBST_JOBQBUF;
990 cb->_aiocb_private.status = cb->aio_nbytes;
996 /* Perform transfer. */
1003 * If we had an error invoking the request, or an error in processing
1004 * the request before we have returned, we process it as an error in
1005 * transfer. Note that such an I/O error is not indicated immediately,
1006 * but is returned using the aio_error mechanism. In this case,
1007 * aio_suspend will return immediately.
1009 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1010 struct aiocb *job = aiocbe->uuaiocb;
1012 aiocbe->uaiocb._aiocb_private.status = 0;
1013 suword(&job->_aiocb_private.status, 0);
1014 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1015 suword(&job->_aiocb_private.error, bp->b_error);
1017 ki->kaio_buffer_finished_count++;
1019 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1020 aiocbe->jobstate = JOBST_JOBBFINISHED;
1021 aiocbe->jobflags |= AIOCBLIST_DONE;
1022 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1023 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1024 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1030 KNOTE(&aiocbe->klist, 0);
1034 ki->kaio_buffer_count--;
1036 lj->lioj_buffer_count--;
1043 * This waits/tests physio completion.
1046 aio_fphysio(struct aiocblist *iocb)
1055 while ((bp->b_flags & B_DONE) == 0) {
1056 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1057 if ((bp->b_flags & B_DONE) == 0) {
1066 /* Release mapping into kernel space. */
1072 /* Check for an error. */
1073 if (bp->b_flags & B_ERROR)
1074 error = bp->b_error;
1079 #endif /* VFS_AIO */
1082 * Wake up aio requests that may be serviceable now.
1085 aio_swake(struct socket *so, struct sockbuf *sb)
1090 struct aiocblist *cb,*cbn;
1092 struct kaioinfo *ki = NULL;
1093 int opcode, wakecount = 0;
1094 struct aioproclist *aiop;
1096 if (sb == &so->so_snd) {
1098 so->so_snd.sb_flags &= ~SB_AIO;
1101 so->so_rcv.sb_flags &= ~SB_AIO;
1104 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1105 cbn = TAILQ_NEXT(cb, list);
1106 if (opcode == cb->uaiocb.aio_lio_opcode) {
1109 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1110 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1111 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1112 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1114 if (cb->jobstate != JOBST_JOBQGLOBAL)
1115 panic("invalid queue value");
1119 while (wakecount--) {
1120 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1121 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1122 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1123 aiop->aioprocflags &= ~AIOP_FREE;
1124 wakeup(aiop->aioproc);
1127 #endif /* VFS_AIO */
1132 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1133 * technique is done in this code.
1136 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1138 struct proc *p = curprpoc;
1139 struct filedesc *fdp;
1145 int opcode, user_opcode;
1146 struct aiocblist *aiocbe;
1147 struct aioproclist *aiop;
1148 struct kaioinfo *ki;
1153 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1154 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1156 aiocbe = zalloc (aiocb_zone);
1158 aiocbe->inputcharge = 0;
1159 aiocbe->outputcharge = 0;
1160 callout_handle_init(&aiocbe->timeouthandle);
1161 SLIST_INIT(&aiocbe->klist);
1163 suword(&job->_aiocb_private.status, -1);
1164 suword(&job->_aiocb_private.error, 0);
1165 suword(&job->_aiocb_private.kernelinfo, -1);
1167 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1169 suword(&job->_aiocb_private.error, error);
1170 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1173 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1174 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1175 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1179 /* Save userspace address of the job info. */
1180 aiocbe->uuaiocb = job;
1182 /* Get the opcode. */
1183 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1184 if (type != LIO_NOP)
1185 aiocbe->uaiocb.aio_lio_opcode = type;
1186 opcode = aiocbe->uaiocb.aio_lio_opcode;
1188 /* Get the fd info for process. */
1192 * Range check file descriptor.
1194 fd = aiocbe->uaiocb.aio_fildes;
1195 if (fd >= fdp->fd_nfiles) {
1196 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1198 suword(&job->_aiocb_private.error, EBADF);
1202 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1203 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1205 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1207 suword(&job->_aiocb_private.error, EBADF);
1212 if (aiocbe->uaiocb.aio_offset == -1LL) {
1216 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1221 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1222 if (jobrefid == LONG_MAX)
1227 if (opcode == LIO_NOP) {
1229 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1231 suword(&job->_aiocb_private.error, 0);
1232 suword(&job->_aiocb_private.status, 0);
1233 suword(&job->_aiocb_private.kernelinfo, 0);
1237 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1239 suword(&job->_aiocb_private.status, 0);
1244 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1245 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1246 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1250 * This method for requesting kevent-based notification won't
1251 * work on the alpha, since we're passing in a pointer
1252 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1253 * based method instead.
1255 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1256 user_opcode == LIO_WRITE)
1259 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1264 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1265 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1266 (kq_fp->f_type != DTYPE_KQUEUE)) {
1270 kq = (struct kqueue *)kq_fp->f_data;
1271 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1272 kev.filter = EVFILT_AIO;
1273 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1274 kev.data = (intptr_t)aiocbe;
1275 error = kqueue_register(kq, &kev, p);
1279 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1281 suword(&job->_aiocb_private.error, error);
1286 suword(&job->_aiocb_private.error, EINPROGRESS);
1287 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1288 aiocbe->userproc = p;
1289 aiocbe->jobflags = 0;
1293 if (fp->f_type == DTYPE_SOCKET) {
1295 * Alternate queueing for socket ops: Reach down into the
1296 * descriptor to get the socket data. Then check to see if the
1297 * socket is ready to be read or written (based on the requested
1300 * If it is not ready for io, then queue the aiocbe on the
1301 * socket, and set the flags so we get a call when sbnotify()
1304 so = (struct socket *)fp->f_data;
1306 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1307 LIO_WRITE) && (!sowriteable(so)))) {
1308 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1309 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1310 if (opcode == LIO_READ)
1311 so->so_rcv.sb_flags |= SB_AIO;
1313 so->so_snd.sb_flags |= SB_AIO;
1314 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1315 ki->kaio_queue_count++;
1324 if ((error = aio_qphysio(p, aiocbe)) == 0)
1327 suword(&job->_aiocb_private.status, 0);
1328 aiocbe->uaiocb._aiocb_private.error = error;
1329 suword(&job->_aiocb_private.error, error);
1333 /* No buffer for daemon I/O. */
1336 ki->kaio_queue_count++;
1338 lj->lioj_queue_count++;
1340 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1341 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1343 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1349 * If we don't have a free AIO process, and we are below our quota, then
1350 * start one. Otherwise, depend on the subsequent I/O completions to
1351 * pick-up this job. If we don't sucessfully create the new process
1352 * (thread) due to resource issues, we return an error for now (EAGAIN),
1353 * which is likely not the correct thing to do.
1357 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1358 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1359 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1360 aiop->aioprocflags &= ~AIOP_FREE;
1361 wakeup(aiop->aioproc);
1362 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1363 ((ki->kaio_active_count + num_aio_resv_start) <
1364 ki->kaio_maxactive_count)) {
1365 num_aio_resv_start++;
1366 if ((error = aio_newproc()) == 0) {
1367 num_aio_resv_start--;
1370 num_aio_resv_start--;
1378 * This routine queues an AIO request, checking for quotas.
1381 aio_aqueue(struct aiocb *job, int type)
1383 struct proc *p = curprpoc;
1384 struct kaioinfo *ki;
1386 if (p->p_aioinfo == NULL)
1387 aio_init_aioinfo(p);
1389 if (num_queue_count >= max_queue_count)
1393 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1396 return _aio_aqueue(job, NULL, type);
1398 #endif /* VFS_AIO */
1401 * Support the aio_return system call, as a side-effect, kernel resources are
1405 aio_return(struct aio_return_args *uap)
1410 struct proc *p = curproc;
1413 struct aiocblist *cb, *ncb;
1415 struct kaioinfo *ki;
1423 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1424 if (jobref == -1 || jobref == 0)
1427 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1428 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1430 if (ujob == cb->uuaiocb) {
1432 cb->uaiocb._aiocb_private.status;
1434 p->p_retval[0] = EFAULT;
1435 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1436 p->p_stats->p_ru.ru_oublock +=
1438 cb->outputcharge = 0;
1439 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1440 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1441 cb->inputcharge = 0;
1448 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1449 ncb = TAILQ_NEXT(cb, plist);
1450 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1453 if (ujob == cb->uuaiocb) {
1455 cb->uaiocb._aiocb_private.status;
1457 p->p_retval[0] = EFAULT;
1465 #endif /* VFS_AIO */
1469 * Allow a process to wakeup when any of the I/O requests are completed.
1472 aio_suspend(struct aio_suspend_args *uap)
1477 struct proc *p = curproc;
1480 struct aiocb *const *cbptr, *cbp;
1481 struct kaioinfo *ki;
1482 struct aiocblist *cb;
1487 struct aiocb **ujoblist;
1489 if (uap->nent > AIO_LISTIO_MAX)
1494 /* Get timespec struct. */
1495 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1498 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1501 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1502 if (itimerfix(&atv))
1504 timo = tvtohz(&atv);
1512 ijoblist = zalloc(aiol_zone);
1513 ujoblist = zalloc(aiol_zone);
1514 cbptr = uap->aiocbp;
1516 for (i = 0; i < uap->nent; i++) {
1517 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1520 ujoblist[njoblist] = cbp;
1521 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1525 if (njoblist == 0) {
1526 zfree(aiol_zone, ijoblist);
1527 zfree(aiol_zone, ujoblist);
1533 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1534 for (i = 0; i < njoblist; i++) {
1536 cb->uaiocb._aiocb_private.kernelinfo) ==
1538 if (ujoblist[i] != cb->uuaiocb)
1540 zfree(aiol_zone, ijoblist);
1541 zfree(aiol_zone, ujoblist);
1548 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1549 TAILQ_NEXT(cb, plist)) {
1550 for (i = 0; i < njoblist; i++) {
1552 cb->uaiocb._aiocb_private.kernelinfo) ==
1555 if (ujoblist[i] != cb->uuaiocb)
1557 zfree(aiol_zone, ijoblist);
1558 zfree(aiol_zone, ujoblist);
1564 ki->kaio_flags |= KAIO_WAKEUP;
1565 error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1568 if (error == ERESTART || error == EINTR) {
1569 zfree(aiol_zone, ijoblist);
1570 zfree(aiol_zone, ujoblist);
1572 } else if (error == EWOULDBLOCK) {
1573 zfree(aiol_zone, ijoblist);
1574 zfree(aiol_zone, ujoblist);
1581 #endif /* VFS_AIO */
1585 * aio_cancel cancels any non-physio aio operations not currently in
1589 aio_cancel(struct aio_cancel_args *uap)
1594 struct proc *p = curproc;
1595 struct kaioinfo *ki;
1596 struct aiocblist *cbe, *cbn;
1598 struct filedesc *fdp;
1607 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1608 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1611 if (fp->f_type == DTYPE_VNODE) {
1612 vp = (struct vnode *)fp->f_data;
1614 if (vn_isdisk(vp,&error)) {
1615 p->p_retval[0] = AIO_NOTCANCELED;
1618 } else if (fp->f_type == DTYPE_SOCKET) {
1619 so = (struct socket *)fp->f_data;
1623 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1624 cbn = TAILQ_NEXT(cbe, list);
1625 if ((uap->aiocbp == NULL) ||
1626 (uap->aiocbp == cbe->uuaiocb) ) {
1629 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1630 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1631 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1632 if (ki->kaio_flags & KAIO_WAKEUP) {
1635 cbe->jobstate = JOBST_JOBFINISHED;
1636 cbe->uaiocb._aiocb_private.status=-1;
1637 cbe->uaiocb._aiocb_private.error=ECANCELED;
1639 /* XXX cancelled, knote? */
1640 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1642 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1649 if ((cancelled) && (uap->aiocbp)) {
1650 p->p_retval[0] = AIO_CANCELED;
1659 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1660 cbn = TAILQ_NEXT(cbe, plist);
1662 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1663 ((uap->aiocbp == NULL ) ||
1664 (uap->aiocbp == cbe->uuaiocb))) {
1666 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1667 TAILQ_REMOVE(&aio_jobs, cbe, list);
1668 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1669 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1672 ki->kaio_queue_finished_count++;
1673 cbe->jobstate = JOBST_JOBFINISHED;
1674 cbe->uaiocb._aiocb_private.status = -1;
1675 cbe->uaiocb._aiocb_private.error = ECANCELED;
1676 /* XXX cancelled, knote? */
1677 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1679 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1688 p->p_retval[0] = AIO_NOTCANCELED;
1692 p->p_retval[0] = AIO_CANCELED;
1695 p->p_retval[0] = AIO_ALLDONE;
1698 #endif /* VFS_AIO */
1702 * aio_error is implemented in the kernel level for compatibility purposes only.
1703 * For a user mode async implementation, it would be best to do it in a userland
1707 aio_error(struct aio_error_args *uap)
1712 struct proc *p = curproc;
1714 struct aiocblist *cb;
1715 struct kaioinfo *ki;
1722 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1723 if ((jobref == -1) || (jobref == 0))
1726 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1727 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1729 p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1736 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1738 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1740 p->p_retval[0] = EINPROGRESS;
1746 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1748 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1750 p->p_retval[0] = EINPROGRESS;
1758 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1760 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1762 p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1768 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1770 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1772 p->p_retval[0] = EINPROGRESS;
1783 status = fuword(&uap->aiocbp->_aiocb_private.status);
1785 return fuword(&uap->aiocbp->_aiocb_private.error);
1788 #endif /* VFS_AIO */
1791 /* syscall - asynchronous read from a file (REALTIME) */
1793 aio_read(struct aio_read_args *uap)
1798 return aio_aqueue(uap->aiocbp, LIO_READ);
1799 #endif /* VFS_AIO */
1802 /* syscall - asynchronous write to a file (REALTIME) */
1804 aio_write(struct aio_write_args *uap)
1809 return aio_aqueue(uap->aiocbp, LIO_WRITE);
1810 #endif /* VFS_AIO */
1813 /* syscall - XXX undocumented */
1815 lio_listio(struct lio_listio_args *uap)
1820 struct proc *p = curproc;
1821 int nent, nentqueued;
1822 struct aiocb *iocb, * const *cbptr;
1823 struct aiocblist *cb;
1824 struct kaioinfo *ki;
1825 struct aio_liojob *lj;
1826 int error, runningcode;
1831 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1835 if (nent > AIO_LISTIO_MAX)
1838 if (p->p_aioinfo == NULL)
1839 aio_init_aioinfo(p);
1841 if ((nent + num_queue_count) > max_queue_count)
1845 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1848 lj = zalloc(aiolio_zone);
1853 lj->lioj_buffer_count = 0;
1854 lj->lioj_buffer_finished_count = 0;
1855 lj->lioj_queue_count = 0;
1856 lj->lioj_queue_finished_count = 0;
1862 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1863 error = copyin(uap->sig, &lj->lioj_signal,
1864 sizeof(lj->lioj_signal));
1866 zfree(aiolio_zone, lj);
1869 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1870 zfree(aiolio_zone, lj);
1873 lj->lioj_flags |= LIOJ_SIGNAL;
1874 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1876 lj->lioj_flags &= ~LIOJ_SIGNAL;
1878 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1880 * Get pointers to the list of I/O requests.
1884 cbptr = uap->acb_list;
1885 for (i = 0; i < uap->nent; i++) {
1886 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1887 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1888 error = _aio_aqueue(iocb, lj, 0);
1897 * If we haven't queued any, then just return error.
1899 if (nentqueued == 0)
1903 * Calculate the appropriate error return.
1909 if (uap->mode == LIO_WAIT) {
1910 int command, found, jobref;
1914 for (i = 0; i < uap->nent; i++) {
1916 * Fetch address of the control buf pointer in
1919 iocb = (struct aiocb *)
1920 (intptr_t)fuword(&cbptr[i]);
1921 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1926 * Fetch the associated command from user space.
1928 command = fuword(&iocb->aio_lio_opcode);
1929 if (command == LIO_NOP) {
1934 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1936 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1937 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1939 if (cb->uaiocb.aio_lio_opcode
1941 p->p_stats->p_ru.ru_oublock
1944 cb->outputcharge = 0;
1945 } else if (cb->uaiocb.aio_lio_opcode
1947 p->p_stats->p_ru.ru_inblock
1949 cb->inputcharge = 0;
1957 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1958 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1968 * If all I/Os have been disposed of, then we can
1971 if (found == nentqueued)
1974 ki->kaio_flags |= KAIO_WAKEUP;
1975 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
1979 else if (error == EWOULDBLOCK)
1985 #endif /* VFS_AIO */
1990 * This is a weird hack so that we can post a signal. It is safe to do so from
1991 * a timeout routine, but *not* from an interrupt routine.
1994 process_signal(void *aioj)
1996 struct aiocblist *aiocbe = aioj;
1997 struct aio_liojob *lj = aiocbe->lio;
1998 struct aiocb *cb = &aiocbe->uaiocb;
2000 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2001 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2002 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2003 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2006 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2007 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2011 * Interrupt handler for physio, performs the necessary process wakeups, and
2015 aio_physwakeup(struct buf *bp)
2017 struct aiocblist *aiocbe;
2019 struct kaioinfo *ki;
2020 struct aio_liojob *lj;
2024 aiocbe = (struct aiocblist *)bp->b_spc;
2028 aiocbe->jobstate = JOBST_JOBBFINISHED;
2029 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2030 aiocbe->uaiocb._aiocb_private.error = 0;
2031 aiocbe->jobflags |= AIOCBLIST_DONE;
2033 if (bp->b_flags & B_ERROR)
2034 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2038 lj->lioj_buffer_finished_count++;
2041 * wakeup/signal if all of the interrupt jobs are done.
2043 if (lj->lioj_buffer_finished_count ==
2044 lj->lioj_buffer_count) {
2046 * Post a signal if it is called for.
2048 if ((lj->lioj_flags &
2049 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2051 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2052 aiocbe->timeouthandle =
2053 timeout(process_signal,
2061 ki->kaio_buffer_finished_count++;
2062 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2063 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2064 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2066 KNOTE(&aiocbe->klist, 0);
2067 /* Do the wakeup. */
2068 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2069 ki->kaio_flags &= ~KAIO_WAKEUP;
2074 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2075 aiocbe->timeouthandle =
2076 timeout(process_signal, aiocbe, 0);
2079 #endif /* VFS_AIO */
2081 /* syscall - wait for the next completion of an aio request */
2083 aio_waitcomplete(struct aio_waitcomplete_args *uap)
2088 struct proc *p = curproc;
2091 struct kaioinfo *ki;
2092 struct aiocblist *cb = NULL;
2095 suword(uap->aiocbp, (int)NULL);
2099 /* Get timespec struct. */
2100 error = copyin(uap->timeout, &ts, sizeof(ts));
2104 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2107 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2108 if (itimerfix(&atv))
2110 timo = tvtohz(&atv);
2118 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2119 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2120 p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2121 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2122 p->p_stats->p_ru.ru_oublock +=
2124 cb->outputcharge = 0;
2125 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2126 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2127 cb->inputcharge = 0;
2130 return cb->uaiocb._aiocb_private.error;
2134 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2136 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2137 p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2139 return cb->uaiocb._aiocb_private.error;
2142 ki->kaio_flags |= KAIO_WAKEUP;
2143 error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2146 if (error == ERESTART)
2150 else if (error == EINTR)
2152 else if (error == EWOULDBLOCK)
2155 #endif /* VFS_AIO */
2160 filt_aioattach(struct knote *kn)
2166 struct filterops aio_filtops =
2167 { 0, filt_aioattach, NULL, NULL };
2170 /* kqueue attach function */
2172 filt_aioattach(struct knote *kn)
2174 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2177 * The aiocbe pointer must be validated before using it, so
2178 * registration is restricted to the kernel; the user cannot
2181 if ((kn->kn_flags & EV_FLAG1) == 0)
2183 kn->kn_flags &= ~EV_FLAG1;
2185 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2190 /* kqueue detach function */
2192 filt_aiodetach(struct knote *kn)
2194 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2196 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2199 /* kqueue filter function */
2202 filt_aio(struct knote *kn, long hint)
2204 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2206 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2207 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2208 aiocbe->jobstate != JOBST_JOBBFINISHED)
2210 kn->kn_flags |= EV_EOF;
2214 struct filterops aio_filtops =
2215 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2216 #endif /* VFS_AIO */