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 1.2 2003/06/17 04:28:41 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 proc *p, 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.
615 aio_daemon(void *uproc)
618 struct aio_liojob *lj;
620 struct aiocblist *aiocbe;
621 struct aioproclist *aiop;
623 struct proc *curcp, *mycp, *userp;
624 struct vmspace *myvm, *tmpvm;
627 * Local copies of curproc (cp) and vmspace (myvm)
630 myvm = mycp->p_vmspace;
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".
666 mycp->p_ucred = crcopy(mycp->p_ucred);
667 mycp->p_ucred->cr_uid = 0;
668 uifree(mycp->p_ucred->cr_uidinfo);
669 mycp->p_ucred->cr_uidinfo = uifind(0);
670 mycp->p_ucred->cr_ngroups = 1;
671 mycp->p_ucred->cr_groups[0] = 1;
673 /* The daemon resides in its own pgrp. */
674 enterpgrp(mycp, mycp->p_pid, 1);
676 /* Mark special process type. */
677 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
680 * Wakeup parent process. (Parent sleeps to keep from blasting away
681 * and creating too many daemons.)
687 * curcp is the current daemon process context.
688 * userp is the current user process context.
693 * Take daemon off of free queue
695 if (aiop->aioprocflags & AIOP_FREE) {
697 TAILQ_REMOVE(&aio_freeproc, aiop, list);
698 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
699 aiop->aioprocflags &= ~AIOP_FREE;
702 aiop->aioprocflags &= ~AIOP_SCHED;
707 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
708 cb = &aiocbe->uaiocb;
709 userp = aiocbe->userproc;
711 aiocbe->jobstate = JOBST_JOBRUNNING;
714 * Connect to process address space for user program.
716 if (userp != curcp) {
718 * Save the current address space that we are
721 tmpvm = mycp->p_vmspace;
724 * Point to the new user address space, and
727 mycp->p_vmspace = userp->p_vmspace;
728 mycp->p_vmspace->vm_refcnt++;
730 /* Activate the new mapping. */
734 * If the old address space wasn't the daemons
735 * own address space, then we need to remove the
736 * daemon's reference from the other process
737 * that it was acting on behalf of.
745 ki = userp->p_aioinfo;
748 /* Account for currently active jobs. */
749 ki->kaio_active_count++;
751 /* Do the I/O function. */
754 /* Decrement the active job count. */
755 ki->kaio_active_count--;
758 * Increment the completion count for wakeup/signal
761 aiocbe->jobflags |= AIOCBLIST_DONE;
762 ki->kaio_queue_finished_count++;
764 lj->lioj_queue_finished_count++;
765 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
766 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
767 ki->kaio_flags &= ~KAIO_WAKEUP;
772 if (lj && (lj->lioj_flags &
773 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
774 if ((lj->lioj_queue_finished_count ==
775 lj->lioj_queue_count) &&
776 (lj->lioj_buffer_finished_count ==
777 lj->lioj_buffer_count)) {
779 lj->lioj_signal.sigev_signo);
786 aiocbe->jobstate = JOBST_JOBFINISHED;
789 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
790 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
792 KNOTE(&aiocbe->klist, 0);
794 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
796 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
799 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
800 psignal(userp, cb->aio_sigevent.sigev_signo);
805 * Disconnect from user address space.
808 /* Get the user address space to disconnect from. */
809 tmpvm = mycp->p_vmspace;
811 /* Get original address space for daemon. */
812 mycp->p_vmspace = myvm;
814 /* Activate the daemon's address space. */
818 printf("AIOD: vmspace problem -- %d\n",
822 /* Remove our vmspace reference. */
829 * If we are the first to be put onto the free queue, wakeup
830 * anyone waiting for a daemon.
833 TAILQ_REMOVE(&aio_activeproc, aiop, list);
834 if (TAILQ_EMPTY(&aio_freeproc))
835 wakeup(&aio_freeproc);
836 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
837 aiop->aioprocflags |= AIOP_FREE;
841 * If daemon is inactive for a long time, allow it to exit,
842 * thereby freeing resources.
844 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
845 PRIBIO, "aiordy", aiod_lifetime)) {
847 if (TAILQ_EMPTY(&aio_jobs)) {
848 if ((aiop->aioprocflags & AIOP_FREE) &&
849 (num_aio_procs > target_aio_procs)) {
850 TAILQ_REMOVE(&aio_freeproc, aiop, list);
852 zfree(aiop_zone, aiop);
855 if (mycp->p_vmspace->vm_refcnt <= 1) {
856 printf("AIOD: bad vm refcnt for"
857 " exiting daemon: %d\n",
858 mycp->p_vmspace->vm_refcnt);
870 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
871 * AIO daemon modifies its environment itself.
880 error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
883 cpu_set_fork_handler(np, aio_daemon, curproc);
886 * Wait until daemon is started, but continue on just in case to
887 * handle error conditions.
889 error = tsleep(np, PZERO, "aiosta", aiod_timeout);
896 * Try the high-performance, low-overhead physio method for eligible
897 * VCHR devices. This method doesn't use an aio helper thread, and
898 * thus has very low overhead.
900 * Assumes that the caller, _aio_aqueue(), has incremented the file
901 * structure's reference count, preventing its deallocation for the
902 * duration of this call.
905 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
913 struct aio_liojob *lj;
917 cb = &aiocbe->uaiocb;
918 fp = aiocbe->fd_file;
920 if (fp->f_type != DTYPE_VNODE)
923 vp = (struct vnode *)fp->f_data;
926 * If its not a disk, we don't want to return a positive error.
927 * It causes the aio code to not fall through to try the thread
928 * way when you're talking to a regular file.
930 if (!vn_isdisk(vp, &error)) {
931 if (error == ENOTBLK)
937 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
941 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
945 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
948 ki->kaio_buffer_count++;
952 lj->lioj_buffer_count++;
954 /* Create and build a buffer header for a transfer. */
955 bp = (struct buf *)getpbuf(NULL);
959 * Get a copy of the kva from the physical buffer.
962 bp->b_dev = vp->v_rdev;
965 bp->b_bcount = cb->aio_nbytes;
966 bp->b_bufsize = cb->aio_nbytes;
967 bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
969 bp->b_iodone = aio_physwakeup;
970 bp->b_saveaddr = bp->b_data;
971 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
972 bp->b_blkno = btodb(cb->aio_offset);
974 /* Bring buffer into kernel space. */
975 if (vmapbuf(bp) < 0) {
982 bp->b_spc = (void *)aiocbe;
983 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
984 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
985 aiocbe->jobstate = JOBST_JOBQBUF;
986 cb->_aiocb_private.status = cb->aio_nbytes;
992 /* Perform transfer. */
999 * If we had an error invoking the request, or an error in processing
1000 * the request before we have returned, we process it as an error in
1001 * transfer. Note that such an I/O error is not indicated immediately,
1002 * but is returned using the aio_error mechanism. In this case,
1003 * aio_suspend will return immediately.
1005 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1006 struct aiocb *job = aiocbe->uuaiocb;
1008 aiocbe->uaiocb._aiocb_private.status = 0;
1009 suword(&job->_aiocb_private.status, 0);
1010 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1011 suword(&job->_aiocb_private.error, bp->b_error);
1013 ki->kaio_buffer_finished_count++;
1015 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1016 aiocbe->jobstate = JOBST_JOBBFINISHED;
1017 aiocbe->jobflags |= AIOCBLIST_DONE;
1018 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1019 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1020 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1026 KNOTE(&aiocbe->klist, 0);
1030 ki->kaio_buffer_count--;
1032 lj->lioj_buffer_count--;
1039 * This waits/tests physio completion.
1042 aio_fphysio(struct aiocblist *iocb)
1051 while ((bp->b_flags & B_DONE) == 0) {
1052 if (tsleep(bp, PRIBIO, "physstr", aiod_timeout)) {
1053 if ((bp->b_flags & B_DONE) == 0) {
1062 /* Release mapping into kernel space. */
1068 /* Check for an error. */
1069 if (bp->b_flags & B_ERROR)
1070 error = bp->b_error;
1075 #endif /* VFS_AIO */
1078 * Wake up aio requests that may be serviceable now.
1081 aio_swake(struct socket *so, struct sockbuf *sb)
1086 struct aiocblist *cb,*cbn;
1088 struct kaioinfo *ki = NULL;
1089 int opcode, wakecount = 0;
1090 struct aioproclist *aiop;
1092 if (sb == &so->so_snd) {
1094 so->so_snd.sb_flags &= ~SB_AIO;
1097 so->so_rcv.sb_flags &= ~SB_AIO;
1100 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1101 cbn = TAILQ_NEXT(cb, list);
1102 if (opcode == cb->uaiocb.aio_lio_opcode) {
1105 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1106 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1107 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1108 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1110 if (cb->jobstate != JOBST_JOBQGLOBAL)
1111 panic("invalid queue value");
1115 while (wakecount--) {
1116 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1117 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1118 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1119 aiop->aioprocflags &= ~AIOP_FREE;
1120 wakeup(aiop->aioproc);
1123 #endif /* VFS_AIO */
1128 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1129 * technique is done in this code.
1132 _aio_aqueue(struct proc *p, struct aiocb *job, struct aio_liojob *lj, int type)
1134 struct filedesc *fdp;
1140 int opcode, user_opcode;
1141 struct aiocblist *aiocbe;
1142 struct aioproclist *aiop;
1143 struct kaioinfo *ki;
1148 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1149 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1151 aiocbe = zalloc (aiocb_zone);
1153 aiocbe->inputcharge = 0;
1154 aiocbe->outputcharge = 0;
1155 callout_handle_init(&aiocbe->timeouthandle);
1156 SLIST_INIT(&aiocbe->klist);
1158 suword(&job->_aiocb_private.status, -1);
1159 suword(&job->_aiocb_private.error, 0);
1160 suword(&job->_aiocb_private.kernelinfo, -1);
1162 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1164 suword(&job->_aiocb_private.error, error);
1165 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1168 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1169 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1170 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1174 /* Save userspace address of the job info. */
1175 aiocbe->uuaiocb = job;
1177 /* Get the opcode. */
1178 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1179 if (type != LIO_NOP)
1180 aiocbe->uaiocb.aio_lio_opcode = type;
1181 opcode = aiocbe->uaiocb.aio_lio_opcode;
1183 /* Get the fd info for process. */
1187 * Range check file descriptor.
1189 fd = aiocbe->uaiocb.aio_fildes;
1190 if (fd >= fdp->fd_nfiles) {
1191 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1193 suword(&job->_aiocb_private.error, EBADF);
1197 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1198 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1200 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1202 suword(&job->_aiocb_private.error, EBADF);
1207 if (aiocbe->uaiocb.aio_offset == -1LL) {
1211 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1216 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1217 if (jobrefid == LONG_MAX)
1222 if (opcode == LIO_NOP) {
1224 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1226 suword(&job->_aiocb_private.error, 0);
1227 suword(&job->_aiocb_private.status, 0);
1228 suword(&job->_aiocb_private.kernelinfo, 0);
1232 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1234 suword(&job->_aiocb_private.status, 0);
1239 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1240 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1241 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1245 * This method for requesting kevent-based notification won't
1246 * work on the alpha, since we're passing in a pointer
1247 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1248 * based method instead.
1250 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1251 user_opcode == LIO_WRITE)
1254 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1259 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1260 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1261 (kq_fp->f_type != DTYPE_KQUEUE)) {
1265 kq = (struct kqueue *)kq_fp->f_data;
1266 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1267 kev.filter = EVFILT_AIO;
1268 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1269 kev.data = (intptr_t)aiocbe;
1270 error = kqueue_register(kq, &kev, p);
1274 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1276 suword(&job->_aiocb_private.error, error);
1281 suword(&job->_aiocb_private.error, EINPROGRESS);
1282 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1283 aiocbe->userproc = p;
1284 aiocbe->jobflags = 0;
1288 if (fp->f_type == DTYPE_SOCKET) {
1290 * Alternate queueing for socket ops: Reach down into the
1291 * descriptor to get the socket data. Then check to see if the
1292 * socket is ready to be read or written (based on the requested
1295 * If it is not ready for io, then queue the aiocbe on the
1296 * socket, and set the flags so we get a call when sbnotify()
1299 so = (struct socket *)fp->f_data;
1301 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1302 LIO_WRITE) && (!sowriteable(so)))) {
1303 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1304 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1305 if (opcode == LIO_READ)
1306 so->so_rcv.sb_flags |= SB_AIO;
1308 so->so_snd.sb_flags |= SB_AIO;
1309 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1310 ki->kaio_queue_count++;
1319 if ((error = aio_qphysio(p, aiocbe)) == 0)
1322 suword(&job->_aiocb_private.status, 0);
1323 aiocbe->uaiocb._aiocb_private.error = error;
1324 suword(&job->_aiocb_private.error, error);
1328 /* No buffer for daemon I/O. */
1331 ki->kaio_queue_count++;
1333 lj->lioj_queue_count++;
1335 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1336 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1338 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1344 * If we don't have a free AIO process, and we are below our quota, then
1345 * start one. Otherwise, depend on the subsequent I/O completions to
1346 * pick-up this job. If we don't sucessfully create the new process
1347 * (thread) due to resource issues, we return an error for now (EAGAIN),
1348 * which is likely not the correct thing to do.
1352 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1353 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1354 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1355 aiop->aioprocflags &= ~AIOP_FREE;
1356 wakeup(aiop->aioproc);
1357 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1358 ((ki->kaio_active_count + num_aio_resv_start) <
1359 ki->kaio_maxactive_count)) {
1360 num_aio_resv_start++;
1361 if ((error = aio_newproc()) == 0) {
1362 num_aio_resv_start--;
1365 num_aio_resv_start--;
1373 * This routine queues an AIO request, checking for quotas.
1376 aio_aqueue(struct proc *p, struct aiocb *job, int type)
1378 struct kaioinfo *ki;
1380 if (p->p_aioinfo == NULL)
1381 aio_init_aioinfo(p);
1383 if (num_queue_count >= max_queue_count)
1387 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1390 return _aio_aqueue(p, job, NULL, type);
1392 #endif /* VFS_AIO */
1395 * Support the aio_return system call, as a side-effect, kernel resources are
1399 aio_return(struct proc *p, struct aio_return_args *uap)
1406 struct aiocblist *cb, *ncb;
1408 struct kaioinfo *ki;
1416 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1417 if (jobref == -1 || jobref == 0)
1420 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1421 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1423 if (ujob == cb->uuaiocb) {
1425 cb->uaiocb._aiocb_private.status;
1427 p->p_retval[0] = EFAULT;
1428 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1429 p->p_stats->p_ru.ru_oublock +=
1431 cb->outputcharge = 0;
1432 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1433 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1434 cb->inputcharge = 0;
1441 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1442 ncb = TAILQ_NEXT(cb, plist);
1443 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1446 if (ujob == cb->uuaiocb) {
1448 cb->uaiocb._aiocb_private.status;
1450 p->p_retval[0] = EFAULT;
1458 #endif /* VFS_AIO */
1462 * Allow a process to wakeup when any of the I/O requests are completed.
1465 aio_suspend(struct proc *p, struct aio_suspend_args *uap)
1472 struct aiocb *const *cbptr, *cbp;
1473 struct kaioinfo *ki;
1474 struct aiocblist *cb;
1479 struct aiocb **ujoblist;
1481 if (uap->nent > AIO_LISTIO_MAX)
1486 /* Get timespec struct. */
1487 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1490 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1493 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1494 if (itimerfix(&atv))
1496 timo = tvtohz(&atv);
1504 ijoblist = zalloc(aiol_zone);
1505 ujoblist = zalloc(aiol_zone);
1506 cbptr = uap->aiocbp;
1508 for (i = 0; i < uap->nent; i++) {
1509 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1512 ujoblist[njoblist] = cbp;
1513 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1517 if (njoblist == 0) {
1518 zfree(aiol_zone, ijoblist);
1519 zfree(aiol_zone, ujoblist);
1525 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1526 for (i = 0; i < njoblist; i++) {
1528 cb->uaiocb._aiocb_private.kernelinfo) ==
1530 if (ujoblist[i] != cb->uuaiocb)
1532 zfree(aiol_zone, ijoblist);
1533 zfree(aiol_zone, ujoblist);
1540 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1541 TAILQ_NEXT(cb, plist)) {
1542 for (i = 0; i < njoblist; i++) {
1544 cb->uaiocb._aiocb_private.kernelinfo) ==
1547 if (ujoblist[i] != cb->uuaiocb)
1549 zfree(aiol_zone, ijoblist);
1550 zfree(aiol_zone, ujoblist);
1556 ki->kaio_flags |= KAIO_WAKEUP;
1557 error = tsleep(p, PRIBIO | PCATCH, "aiospn", timo);
1560 if (error == ERESTART || error == EINTR) {
1561 zfree(aiol_zone, ijoblist);
1562 zfree(aiol_zone, ujoblist);
1564 } else if (error == EWOULDBLOCK) {
1565 zfree(aiol_zone, ijoblist);
1566 zfree(aiol_zone, ujoblist);
1573 #endif /* VFS_AIO */
1577 * aio_cancel cancels any non-physio aio operations not currently in
1581 aio_cancel(struct proc *p, struct aio_cancel_args *uap)
1586 struct kaioinfo *ki;
1587 struct aiocblist *cbe, *cbn;
1589 struct filedesc *fdp;
1598 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1599 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1602 if (fp->f_type == DTYPE_VNODE) {
1603 vp = (struct vnode *)fp->f_data;
1605 if (vn_isdisk(vp,&error)) {
1606 p->p_retval[0] = AIO_NOTCANCELED;
1609 } else if (fp->f_type == DTYPE_SOCKET) {
1610 so = (struct socket *)fp->f_data;
1614 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1615 cbn = TAILQ_NEXT(cbe, list);
1616 if ((uap->aiocbp == NULL) ||
1617 (uap->aiocbp == cbe->uuaiocb) ) {
1620 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1621 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1622 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1623 if (ki->kaio_flags & KAIO_WAKEUP) {
1626 cbe->jobstate = JOBST_JOBFINISHED;
1627 cbe->uaiocb._aiocb_private.status=-1;
1628 cbe->uaiocb._aiocb_private.error=ECANCELED;
1630 /* XXX cancelled, knote? */
1631 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1633 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1640 if ((cancelled) && (uap->aiocbp)) {
1641 p->p_retval[0] = AIO_CANCELED;
1650 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1651 cbn = TAILQ_NEXT(cbe, plist);
1653 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1654 ((uap->aiocbp == NULL ) ||
1655 (uap->aiocbp == cbe->uuaiocb))) {
1657 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1658 TAILQ_REMOVE(&aio_jobs, cbe, list);
1659 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1660 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1663 ki->kaio_queue_finished_count++;
1664 cbe->jobstate = JOBST_JOBFINISHED;
1665 cbe->uaiocb._aiocb_private.status = -1;
1666 cbe->uaiocb._aiocb_private.error = ECANCELED;
1667 /* XXX cancelled, knote? */
1668 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1670 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1679 p->p_retval[0] = AIO_NOTCANCELED;
1683 p->p_retval[0] = AIO_CANCELED;
1686 p->p_retval[0] = AIO_ALLDONE;
1689 #endif /* VFS_AIO */
1693 * aio_error is implemented in the kernel level for compatibility purposes only.
1694 * For a user mode async implementation, it would be best to do it in a userland
1698 aio_error(struct proc *p, struct aio_error_args *uap)
1704 struct aiocblist *cb;
1705 struct kaioinfo *ki;
1712 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1713 if ((jobref == -1) || (jobref == 0))
1716 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1717 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1719 p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1726 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1728 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1730 p->p_retval[0] = EINPROGRESS;
1736 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1738 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1740 p->p_retval[0] = EINPROGRESS;
1748 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1750 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1752 p->p_retval[0] = cb->uaiocb._aiocb_private.error;
1758 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1760 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1762 p->p_retval[0] = EINPROGRESS;
1773 status = fuword(&uap->aiocbp->_aiocb_private.status);
1775 return fuword(&uap->aiocbp->_aiocb_private.error);
1778 #endif /* VFS_AIO */
1781 /* syscall - asynchronous read from a file (REALTIME) */
1783 aio_read(struct proc *p, struct aio_read_args *uap)
1788 return aio_aqueue(p, uap->aiocbp, LIO_READ);
1789 #endif /* VFS_AIO */
1792 /* syscall - asynchronous write to a file (REALTIME) */
1794 aio_write(struct proc *p, struct aio_write_args *uap)
1799 return aio_aqueue(p, uap->aiocbp, LIO_WRITE);
1800 #endif /* VFS_AIO */
1803 /* syscall - XXX undocumented */
1805 lio_listio(struct proc *p, struct lio_listio_args *uap)
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;
1820 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1824 if (nent > AIO_LISTIO_MAX)
1827 if (p->p_aioinfo == NULL)
1828 aio_init_aioinfo(p);
1830 if ((nent + num_queue_count) > max_queue_count)
1834 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1837 lj = zalloc(aiolio_zone);
1842 lj->lioj_buffer_count = 0;
1843 lj->lioj_buffer_finished_count = 0;
1844 lj->lioj_queue_count = 0;
1845 lj->lioj_queue_finished_count = 0;
1851 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1852 error = copyin(uap->sig, &lj->lioj_signal,
1853 sizeof(lj->lioj_signal));
1855 zfree(aiolio_zone, lj);
1858 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1859 zfree(aiolio_zone, lj);
1862 lj->lioj_flags |= LIOJ_SIGNAL;
1863 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1865 lj->lioj_flags &= ~LIOJ_SIGNAL;
1867 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1869 * Get pointers to the list of I/O requests.
1873 cbptr = uap->acb_list;
1874 for (i = 0; i < uap->nent; i++) {
1875 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1876 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1877 error = _aio_aqueue(p, iocb, lj, 0);
1886 * If we haven't queued any, then just return error.
1888 if (nentqueued == 0)
1892 * Calculate the appropriate error return.
1898 if (uap->mode == LIO_WAIT) {
1899 int command, found, jobref;
1903 for (i = 0; i < uap->nent; i++) {
1905 * Fetch address of the control buf pointer in
1908 iocb = (struct aiocb *)
1909 (intptr_t)fuword(&cbptr[i]);
1910 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1915 * Fetch the associated command from user space.
1917 command = fuword(&iocb->aio_lio_opcode);
1918 if (command == LIO_NOP) {
1923 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1925 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1926 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1928 if (cb->uaiocb.aio_lio_opcode
1930 p->p_stats->p_ru.ru_oublock
1933 cb->outputcharge = 0;
1934 } else if (cb->uaiocb.aio_lio_opcode
1936 p->p_stats->p_ru.ru_inblock
1938 cb->inputcharge = 0;
1946 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1947 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1957 * If all I/Os have been disposed of, then we can
1960 if (found == nentqueued)
1963 ki->kaio_flags |= KAIO_WAKEUP;
1964 error = tsleep(p, PRIBIO | PCATCH, "aiospn", 0);
1968 else if (error == EWOULDBLOCK)
1974 #endif /* VFS_AIO */
1979 * This is a weird hack so that we can post a signal. It is safe to do so from
1980 * a timeout routine, but *not* from an interrupt routine.
1983 process_signal(void *aioj)
1985 struct aiocblist *aiocbe = aioj;
1986 struct aio_liojob *lj = aiocbe->lio;
1987 struct aiocb *cb = &aiocbe->uaiocb;
1989 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
1990 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
1991 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
1992 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
1995 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
1996 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2000 * Interrupt handler for physio, performs the necessary process wakeups, and
2004 aio_physwakeup(struct buf *bp)
2006 struct aiocblist *aiocbe;
2008 struct kaioinfo *ki;
2009 struct aio_liojob *lj;
2013 aiocbe = (struct aiocblist *)bp->b_spc;
2017 aiocbe->jobstate = JOBST_JOBBFINISHED;
2018 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2019 aiocbe->uaiocb._aiocb_private.error = 0;
2020 aiocbe->jobflags |= AIOCBLIST_DONE;
2022 if (bp->b_flags & B_ERROR)
2023 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2027 lj->lioj_buffer_finished_count++;
2030 * wakeup/signal if all of the interrupt jobs are done.
2032 if (lj->lioj_buffer_finished_count ==
2033 lj->lioj_buffer_count) {
2035 * Post a signal if it is called for.
2037 if ((lj->lioj_flags &
2038 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2040 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2041 aiocbe->timeouthandle =
2042 timeout(process_signal,
2050 ki->kaio_buffer_finished_count++;
2051 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2052 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2053 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2055 KNOTE(&aiocbe->klist, 0);
2056 /* Do the wakeup. */
2057 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2058 ki->kaio_flags &= ~KAIO_WAKEUP;
2063 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2064 aiocbe->timeouthandle =
2065 timeout(process_signal, aiocbe, 0);
2068 #endif /* VFS_AIO */
2070 /* syscall - wait for the next completion of an aio request */
2072 aio_waitcomplete(struct proc *p, struct aio_waitcomplete_args *uap)
2079 struct kaioinfo *ki;
2080 struct aiocblist *cb = NULL;
2083 suword(uap->aiocbp, (int)NULL);
2087 /* Get timespec struct. */
2088 error = copyin(uap->timeout, &ts, sizeof(ts));
2092 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2095 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2096 if (itimerfix(&atv))
2098 timo = tvtohz(&atv);
2106 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2107 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2108 p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2109 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2110 p->p_stats->p_ru.ru_oublock +=
2112 cb->outputcharge = 0;
2113 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2114 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2115 cb->inputcharge = 0;
2118 return cb->uaiocb._aiocb_private.error;
2122 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2124 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2125 p->p_retval[0] = cb->uaiocb._aiocb_private.status;
2127 return cb->uaiocb._aiocb_private.error;
2130 ki->kaio_flags |= KAIO_WAKEUP;
2131 error = tsleep(p, PRIBIO | PCATCH, "aiowc", timo);
2134 if (error == ERESTART)
2138 else if (error == EINTR)
2140 else if (error == EWOULDBLOCK)
2143 #endif /* VFS_AIO */
2148 filt_aioattach(struct knote *kn)
2154 struct filterops aio_filtops =
2155 { 0, filt_aioattach, NULL, NULL };
2158 /* kqueue attach function */
2160 filt_aioattach(struct knote *kn)
2162 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2165 * The aiocbe pointer must be validated before using it, so
2166 * registration is restricted to the kernel; the user cannot
2169 if ((kn->kn_flags & EV_FLAG1) == 0)
2171 kn->kn_flags &= ~EV_FLAG1;
2173 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2178 /* kqueue detach function */
2180 filt_aiodetach(struct knote *kn)
2182 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2184 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2187 /* kqueue filter function */
2190 filt_aio(struct knote *kn, long hint)
2192 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2194 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2195 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2196 aiocbe->jobstate != JOBST_JOBBFINISHED)
2198 kn->kn_flags |= EV_EOF;
2202 struct filterops aio_filtops =
2203 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2204 #endif /* VFS_AIO */