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 1.35 2007/02/03 17:05:58 corecode Exp $
21 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
24 #include <sys/param.h>
25 #include <sys/systm.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
33 #include <sys/unistd.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
42 #include <sys/event.h>
45 #include <vm/vm_extern.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
50 #include <sys/file2.h>
52 #include <sys/thread2.h>
54 #include <machine/limits.h>
55 #include "opt_vfs_aio.h"
60 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
65 #define JOBST_NULL 0x0
66 #define JOBST_JOBQGLOBAL 0x2
67 #define JOBST_JOBRUNNING 0x3
68 #define JOBST_JOBFINISHED 0x4
69 #define JOBST_JOBQBUF 0x5
70 #define JOBST_JOBBFINISHED 0x6
72 #ifndef MAX_AIO_PER_PROC
73 #define MAX_AIO_PER_PROC 32
76 #ifndef MAX_AIO_QUEUE_PER_PROC
77 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
81 #define MAX_AIO_PROCS 32
85 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
88 #ifndef TARGET_AIO_PROCS
89 #define TARGET_AIO_PROCS 4
93 #define MAX_BUF_AIO 16
96 #ifndef AIOD_TIMEOUT_DEFAULT
97 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
100 #ifndef AIOD_LIFETIME_DEFAULT
101 #define AIOD_LIFETIME_DEFAULT (30 * hz)
104 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
106 static int max_aio_procs = MAX_AIO_PROCS;
107 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
108 CTLFLAG_RW, &max_aio_procs, 0,
109 "Maximum number of kernel threads to use for handling async IO");
111 static int num_aio_procs = 0;
112 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
113 CTLFLAG_RD, &num_aio_procs, 0,
114 "Number of presently active kernel threads for async IO");
117 * The code will adjust the actual number of AIO processes towards this
118 * number when it gets a chance.
120 static int target_aio_procs = TARGET_AIO_PROCS;
121 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
122 0, "Preferred number of ready kernel threads for async IO");
124 static int max_queue_count = MAX_AIO_QUEUE;
125 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
126 "Maximum number of aio requests to queue, globally");
128 static int num_queue_count = 0;
129 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
130 "Number of queued aio requests");
132 static int num_buf_aio = 0;
133 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
134 "Number of aio requests presently handled by the buf subsystem");
136 /* Number of async I/O thread in the process of being started */
137 /* XXX This should be local to _aio_aqueue() */
138 static int num_aio_resv_start = 0;
140 static int aiod_timeout;
141 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
142 "Timeout value for synchronous aio operations");
144 static int aiod_lifetime;
145 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
146 "Maximum lifetime for idle aiod");
148 static int max_aio_per_proc = MAX_AIO_PER_PROC;
149 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
150 0, "Maximum active aio requests per process (stored in the process)");
152 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
153 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
154 &max_aio_queue_per_proc, 0,
155 "Maximum queued aio requests per process (stored in the process)");
157 static int max_buf_aio = MAX_BUF_AIO;
158 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
159 "Maximum buf aio requests per process (stored in the process)");
164 #define AIOP_FREE 0x1 /* proc on free queue */
165 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
168 int aioprocflags; /* AIO proc flags */
169 TAILQ_ENTRY(aioproclist) list; /* List of processes */
170 struct proc *aioproc; /* The AIO thread */
174 * data-structure for lio signal management
178 int lioj_buffer_count;
179 int lioj_buffer_finished_count;
180 int lioj_queue_count;
181 int lioj_queue_finished_count;
182 struct sigevent lioj_signal; /* signal on all I/O done */
183 TAILQ_ENTRY(aio_liojob) lioj_list;
184 struct kaioinfo *lioj_ki;
186 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
187 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
190 * per process aio data structure
193 int kaio_flags; /* per process kaio flags */
194 int kaio_maxactive_count; /* maximum number of AIOs */
195 int kaio_active_count; /* number of currently used AIOs */
196 int kaio_qallowed_count; /* maxiumu size of AIO queue */
197 int kaio_queue_count; /* size of AIO queue */
198 int kaio_ballowed_count; /* maximum number of buffers */
199 int kaio_queue_finished_count; /* number of daemon jobs finished */
200 int kaio_buffer_count; /* number of physio buffers */
201 int kaio_buffer_finished_count; /* count of I/O done */
202 struct proc *kaio_p; /* process that uses this kaio block */
203 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
204 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
205 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
206 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
207 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
208 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
211 #define KAIO_RUNDOWN 0x1 /* process is being run down */
212 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
214 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
215 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
216 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
217 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
219 static void aio_init_aioinfo(struct proc *p);
220 static void aio_onceonly(void *);
221 static int aio_free_entry(struct aiocblist *aiocbe);
222 static void aio_process(struct aiocblist *aiocbe);
223 static int aio_newproc(void);
224 static int aio_aqueue(struct aiocb *job, int type);
225 static void aio_physwakeup(struct bio *bio);
226 static int aio_fphysio(struct aiocblist *aiocbe);
227 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
228 static void aio_daemon(void *uproc);
229 static void process_signal(void *aioj);
231 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
235 * kaio Per process async io info
236 * aiop async io thread data
237 * aiocb async io jobs
238 * aiol list io job pointer - internal to aio_suspend XXX
239 * aiolio list io jobs
241 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
244 * Startup initialization
247 aio_onceonly(void *na)
249 TAILQ_INIT(&aio_freeproc);
250 TAILQ_INIT(&aio_activeproc);
251 TAILQ_INIT(&aio_jobs);
252 TAILQ_INIT(&aio_bufjobs);
253 TAILQ_INIT(&aio_freejobs);
254 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
255 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
256 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
257 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
258 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
259 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
260 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
265 * Init the per-process aioinfo structure. The aioinfo limits are set
266 * per-process for user limit (resource) management.
269 aio_init_aioinfo(struct proc *p)
272 if (p->p_aioinfo == NULL) {
273 ki = zalloc(kaio_zone);
276 ki->kaio_maxactive_count = max_aio_per_proc;
277 ki->kaio_active_count = 0;
278 ki->kaio_qallowed_count = max_aio_queue_per_proc;
279 ki->kaio_queue_count = 0;
280 ki->kaio_ballowed_count = max_buf_aio;
281 ki->kaio_buffer_count = 0;
282 ki->kaio_buffer_finished_count = 0;
284 TAILQ_INIT(&ki->kaio_jobdone);
285 TAILQ_INIT(&ki->kaio_jobqueue);
286 TAILQ_INIT(&ki->kaio_bufdone);
287 TAILQ_INIT(&ki->kaio_bufqueue);
288 TAILQ_INIT(&ki->kaio_liojoblist);
289 TAILQ_INIT(&ki->kaio_sockqueue);
292 while (num_aio_procs < target_aio_procs)
297 * Free a job entry. Wait for completion if it is currently active, but don't
298 * delay forever. If we delay, we return a flag that says that we have to
299 * restart the queue scan.
302 aio_free_entry(struct aiocblist *aiocbe)
305 struct aio_liojob *lj;
309 if (aiocbe->jobstate == JOBST_NULL)
310 panic("aio_free_entry: freeing already free job");
312 p = aiocbe->userproc;
316 panic("aio_free_entry: missing p->p_aioinfo");
318 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
319 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
320 tsleep(aiocbe, 0, "jobwai", 0);
322 if (aiocbe->bp == NULL) {
323 if (ki->kaio_queue_count <= 0)
324 panic("aio_free_entry: process queue size <= 0");
325 if (num_queue_count <= 0)
326 panic("aio_free_entry: system wide queue size <= 0");
329 lj->lioj_queue_count--;
330 if (aiocbe->jobflags & AIOCBLIST_DONE)
331 lj->lioj_queue_finished_count--;
333 ki->kaio_queue_count--;
334 if (aiocbe->jobflags & AIOCBLIST_DONE)
335 ki->kaio_queue_finished_count--;
339 lj->lioj_buffer_count--;
340 if (aiocbe->jobflags & AIOCBLIST_DONE)
341 lj->lioj_buffer_finished_count--;
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 ki->kaio_buffer_finished_count--;
345 ki->kaio_buffer_count--;
349 /* aiocbe is going away, we need to destroy any knotes */
350 /* XXX lwp knote wants a thread, but only cares about the process */
351 knote_remove(FIRST_LWP_IN_PROC(p)->lwp_thread, &aiocbe->klist);
353 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
354 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
355 ki->kaio_flags &= ~KAIO_WAKEUP;
359 if (aiocbe->jobstate == JOBST_JOBQBUF) {
360 if ((error = aio_fphysio(aiocbe)) != 0)
362 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
363 panic("aio_free_entry: invalid physio finish-up state");
365 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
367 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
369 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
370 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
372 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
373 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
374 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
376 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
379 vunmapbuf(aiocbe->bp);
380 relpbuf(aiocbe->bp, NULL);
384 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
385 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
386 zfree(aiolio_zone, lj);
388 aiocbe->jobstate = JOBST_NULL;
389 callout_stop(&aiocbe->timeout);
390 fdrop(aiocbe->fd_file);
391 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
397 * Rundown the jobs for a given process.
400 aio_proc_rundown(struct proc *p)
406 struct aio_liojob *lj, *ljn;
407 struct aiocblist *aiocbe, *aiocbn;
415 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
416 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
417 ki->kaio_buffer_finished_count)) {
418 ki->kaio_flags |= KAIO_RUNDOWN;
419 if (tsleep(p, 0, "kaiowt", aiod_timeout))
424 * Move any aio ops that are waiting on socket I/O to the normal job
425 * queues so they are cleaned up with any others.
428 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
430 aiocbn = TAILQ_NEXT(aiocbe, plist);
431 fp = aiocbe->fd_file;
433 so = (struct socket *)fp->f_data;
434 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
435 if (TAILQ_EMPTY(&so->so_aiojobq)) {
436 so->so_snd.sb_flags &= ~SB_AIO;
437 so->so_rcv.sb_flags &= ~SB_AIO;
440 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
441 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
442 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
447 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
448 aiocbn = TAILQ_NEXT(aiocbe, plist);
449 if (aio_free_entry(aiocbe))
454 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
456 aiocbn = TAILQ_NEXT(aiocbe, plist);
457 if (aio_free_entry(aiocbe))
463 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
464 ki->kaio_flags |= KAIO_WAKEUP;
465 tsleep(p, 0, "aioprn", 0);
473 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
474 aiocbn = TAILQ_NEXT(aiocbe, plist);
475 if (aio_free_entry(aiocbe)) {
483 * If we've slept, jobs might have moved from one queue to another.
484 * Retry rundown if we didn't manage to empty the queues.
486 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
487 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
488 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
489 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
492 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
493 ljn = TAILQ_NEXT(lj, lioj_list);
494 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
496 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
497 zfree(aiolio_zone, lj);
500 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
501 "QF:%d\n", lj->lioj_buffer_count,
502 lj->lioj_buffer_finished_count,
503 lj->lioj_queue_count,
504 lj->lioj_queue_finished_count);
509 zfree(kaio_zone, ki);
516 * Select a job to run (called by an AIO daemon).
518 static struct aiocblist *
519 aio_selectjob(struct aioproclist *aiop)
521 struct aiocblist *aiocbe;
526 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
527 TAILQ_NEXT(aiocbe, list)) {
528 userp = aiocbe->userproc;
529 ki = userp->p_aioinfo;
531 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
532 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
543 * The AIO processing activity. This is the code that does the I/O request for
544 * the non-physio version of the operations. The normal vn operations are used,
545 * and this code should work in all instances for every type of file, including
546 * pipes, sockets, fifos, and regular files.
549 aio_process(struct aiocblist *aiocbe)
558 int oublock_st, oublock_end;
559 int inblock_st, inblock_end;
562 cb = &aiocbe->uaiocb;
563 fp = aiocbe->fd_file;
565 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
566 aiov.iov_len = cb->aio_nbytes;
568 auio.uio_iov = &aiov;
570 auio.uio_offset = cb->aio_offset;
571 auio.uio_resid = cb->aio_nbytes;
572 cnt = cb->aio_nbytes;
573 auio.uio_segflg = UIO_USERSPACE;
576 inblock_st = mytd->td_lwp->lwp_ru.ru_inblock;
577 oublock_st = mytd->td_lwp->lwp_ru.ru_oublock;
579 * _aio_aqueue() acquires a reference to the file that is
580 * released in aio_free_entry().
582 if (cb->aio_lio_opcode == LIO_READ) {
583 auio.uio_rw = UIO_READ;
584 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET);
586 auio.uio_rw = UIO_WRITE;
587 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET);
589 inblock_end = mytd->td_lwp->lwp_ru.ru_inblock;
590 oublock_end = mytd->td_lwp->lwp_ru.ru_oublock;
592 aiocbe->inputcharge = inblock_end - inblock_st;
593 aiocbe->outputcharge = oublock_end - oublock_st;
595 if ((error) && (auio.uio_resid != cnt)) {
596 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
598 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
599 ksignal(aiocbe->userproc, SIGPIPE);
602 cnt -= auio.uio_resid;
603 cb->_aiocb_private.error = error;
604 cb->_aiocb_private.status = cnt;
608 * The AIO daemon, most of the actual work is done in aio_process,
609 * but the setup (and address space mgmt) is done in this routine.
611 * The MP lock is held on entry.
614 aio_daemon(void *uproc)
616 struct aio_liojob *lj;
618 struct aiocblist *aiocbe;
619 struct aioproclist *aiop;
621 struct proc *curcp, *mycp, *userp;
622 struct vmspace *myvm, *tmpvm;
626 * Local copies of curproc (cp) and vmspace (myvm)
629 myvm = mycp->p_vmspace;
631 if (mycp->p_textvp) {
632 vrele(mycp->p_textvp);
633 mycp->p_textvp = NULL;
637 * Allocate and ready the aio control info. There is one aiop structure
640 aiop = zalloc(aiop_zone);
641 aiop->aioproc = mycp;
642 aiop->aioprocflags |= AIOP_FREE;
647 * Place thread (lightweight process) onto the AIO free thread list.
649 if (TAILQ_EMPTY(&aio_freeproc))
650 wakeup(&aio_freeproc);
651 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
655 /* Make up a name for the daemon. */
656 strcpy(mycp->p_comm, "aiod");
659 * Get rid of our current filedescriptors. AIOD's don't need any
660 * filedescriptors, except as temporarily inherited from the client.
661 * 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.)
685 * curcp is the current daemon process context.
686 * userp is the current user process context.
691 * Take daemon off of free queue
693 if (aiop->aioprocflags & AIOP_FREE) {
695 TAILQ_REMOVE(&aio_freeproc, aiop, list);
696 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
697 aiop->aioprocflags &= ~AIOP_FREE;
700 aiop->aioprocflags &= ~AIOP_SCHED;
705 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
706 cb = &aiocbe->uaiocb;
707 userp = aiocbe->userproc;
709 aiocbe->jobstate = JOBST_JOBRUNNING;
712 * Connect to process address space for user program.
714 if (userp != curcp) {
716 * Save the current address space that we are
719 tmpvm = mycp->p_vmspace;
722 * Point to the new user address space, and
725 mycp->p_vmspace = userp->p_vmspace;
726 mycp->p_vmspace->vm_refcnt++;
728 /* Activate the new mapping. */
732 * If the old address space wasn't the daemons
733 * own address space, then we need to remove the
734 * daemon's reference from the other process
735 * that it was acting on behalf of.
743 ki = userp->p_aioinfo;
746 /* Account for currently active jobs. */
747 ki->kaio_active_count++;
749 /* Do the I/O function. */
752 /* Decrement the active job count. */
753 ki->kaio_active_count--;
756 * Increment the completion count for wakeup/signal
759 aiocbe->jobflags |= AIOCBLIST_DONE;
760 ki->kaio_queue_finished_count++;
762 lj->lioj_queue_finished_count++;
763 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
764 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
765 ki->kaio_flags &= ~KAIO_WAKEUP;
770 if (lj && (lj->lioj_flags &
771 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
772 if ((lj->lioj_queue_finished_count ==
773 lj->lioj_queue_count) &&
774 (lj->lioj_buffer_finished_count ==
775 lj->lioj_buffer_count)) {
777 lj->lioj_signal.sigev_signo);
784 aiocbe->jobstate = JOBST_JOBFINISHED;
787 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
788 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
790 KNOTE(&aiocbe->klist, 0);
792 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
794 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
797 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
798 ksignal(userp, cb->aio_sigevent.sigev_signo);
803 * Disconnect from user address space.
806 /* Get the user address space to disconnect from. */
807 tmpvm = mycp->p_vmspace;
809 /* Get original address space for daemon. */
810 mycp->p_vmspace = myvm;
812 /* Activate the daemon's address space. */
816 kprintf("AIOD: vmspace problem -- %d\n",
820 /* Remove our vmspace reference. */
827 * If we are the first to be put onto the free queue, wakeup
828 * anyone waiting for a daemon.
831 TAILQ_REMOVE(&aio_activeproc, aiop, list);
832 if (TAILQ_EMPTY(&aio_freeproc))
833 wakeup(&aio_freeproc);
834 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
835 aiop->aioprocflags |= AIOP_FREE;
839 * If daemon is inactive for a long time, allow it to exit,
840 * thereby freeing resources.
842 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
843 0, "aiordy", aiod_lifetime)) {
845 if (TAILQ_EMPTY(&aio_jobs)) {
846 if ((aiop->aioprocflags & AIOP_FREE) &&
847 (num_aio_procs > target_aio_procs)) {
848 TAILQ_REMOVE(&aio_freeproc, aiop, list);
850 zfree(aiop_zone, aiop);
853 if (mycp->p_vmspace->vm_refcnt <= 1) {
854 kprintf("AIOD: bad vm refcnt for"
855 " exiting daemon: %d\n",
856 mycp->p_vmspace->vm_refcnt);
868 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
869 * AIO daemon modifies its environment itself.
875 struct lwp *lp, *nlp;
879 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
882 nlp = ONLY_LWP_IN_PROC(np);
883 cpu_set_fork_handler(nlp, aio_daemon, curproc);
884 start_forked_proc(lp, np);
887 * Wait until daemon is started, but continue on just in case to
888 * handle error conditions.
890 error = tsleep(np, 0, "aiosta", aiod_timeout);
897 * Try the high-performance, low-overhead physio method for eligible
898 * VCHR devices. This method doesn't use an aio helper thread, and
899 * thus has very low overhead.
901 * Assumes that the caller, _aio_aqueue(), has incremented the file
902 * structure's reference count, preventing its deallocation for the
903 * duration of this call.
906 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
914 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. */
959 * Get a copy of the kva from the physical buffer.
961 bp->b_bio1.bio_caller_info1.ptr = p;
964 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
965 BUF_CMD_WRITE : BUF_CMD_READ;
966 bp->b_bio1.bio_done = aio_physwakeup;
967 bp->b_bio1.bio_offset = cb->aio_offset;
969 /* Bring buffer into kernel space. */
970 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
978 bp->b_bio1.bio_caller_info2.ptr = aiocbe;
979 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
980 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
981 aiocbe->jobstate = JOBST_JOBQBUF;
982 cb->_aiocb_private.status = cb->aio_nbytes;
988 /* Perform transfer. */
989 dev_dstrategy(vp->v_rdev, &bp->b_bio1);
995 * If we had an error invoking the request, or an error in processing
996 * the request before we have returned, we process it as an error in
997 * transfer. Note that such an I/O error is not indicated immediately,
998 * but is returned using the aio_error mechanism. In this case,
999 * aio_suspend will return immediately.
1001 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1002 struct aiocb *job = aiocbe->uuaiocb;
1004 aiocbe->uaiocb._aiocb_private.status = 0;
1005 suword(&job->_aiocb_private.status, 0);
1006 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1007 suword(&job->_aiocb_private.error, bp->b_error);
1009 ki->kaio_buffer_finished_count++;
1011 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1012 aiocbe->jobstate = JOBST_JOBBFINISHED;
1013 aiocbe->jobflags |= AIOCBLIST_DONE;
1014 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1015 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1016 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1022 KNOTE(&aiocbe->klist, 0);
1026 ki->kaio_buffer_count--;
1028 lj->lioj_buffer_count--;
1035 * This waits/tests physio completion.
1038 aio_fphysio(struct aiocblist *iocb)
1046 while (bp->b_cmd != BUF_CMD_DONE) {
1047 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
1048 if (bp->b_cmd != BUF_CMD_DONE) {
1058 /* Release mapping into kernel space. */
1064 /* Check for an error. */
1065 if (bp->b_flags & B_ERROR)
1066 error = bp->b_error;
1071 #endif /* VFS_AIO */
1074 * Wake up aio requests that may be serviceable now.
1077 aio_swake(struct socket *so, struct sockbuf *sb)
1082 struct aiocblist *cb,*cbn;
1084 struct kaioinfo *ki = NULL;
1085 int opcode, wakecount = 0;
1086 struct aioproclist *aiop;
1088 if (sb == &so->so_snd) {
1090 so->so_snd.sb_flags &= ~SB_AIO;
1093 so->so_rcv.sb_flags &= ~SB_AIO;
1096 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1097 cbn = TAILQ_NEXT(cb, list);
1098 if (opcode == cb->uaiocb.aio_lio_opcode) {
1101 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1102 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1103 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1104 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1106 if (cb->jobstate != JOBST_JOBQGLOBAL)
1107 panic("invalid queue value");
1111 while (wakecount--) {
1112 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1113 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1114 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1115 aiop->aioprocflags &= ~AIOP_FREE;
1116 wakeup(aiop->aioproc);
1119 #endif /* VFS_AIO */
1124 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1125 * technique is done in this code.
1128 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1130 struct proc *p = curproc;
1131 struct filedesc *fdp;
1136 int opcode, user_opcode;
1137 struct aiocblist *aiocbe;
1138 struct aioproclist *aiop;
1139 struct kaioinfo *ki;
1144 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1145 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1147 aiocbe = zalloc (aiocb_zone);
1149 aiocbe->inputcharge = 0;
1150 aiocbe->outputcharge = 0;
1151 callout_init(&aiocbe->timeout);
1152 SLIST_INIT(&aiocbe->klist);
1154 suword(&job->_aiocb_private.status, -1);
1155 suword(&job->_aiocb_private.error, 0);
1156 suword(&job->_aiocb_private.kernelinfo, -1);
1158 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1160 suword(&job->_aiocb_private.error, error);
1161 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1164 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1165 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1166 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1170 /* Save userspace address of the job info. */
1171 aiocbe->uuaiocb = job;
1173 /* Get the opcode. */
1174 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1175 if (type != LIO_NOP)
1176 aiocbe->uaiocb.aio_lio_opcode = type;
1177 opcode = aiocbe->uaiocb.aio_lio_opcode;
1179 /* Get the fd info for process. */
1183 * Range check file descriptor.
1185 fd = aiocbe->uaiocb.aio_fildes;
1186 if (fd >= fdp->fd_nfiles) {
1187 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1189 suword(&job->_aiocb_private.error, EBADF);
1193 fp = aiocbe->fd_file = fdp->fd_files[fd].fp;
1194 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1196 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1198 suword(&job->_aiocb_private.error, EBADF);
1203 if (aiocbe->uaiocb.aio_offset == -1LL) {
1207 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1212 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1213 if (jobrefid == LONG_MAX)
1218 if (opcode == LIO_NOP) {
1220 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1222 suword(&job->_aiocb_private.error, 0);
1223 suword(&job->_aiocb_private.status, 0);
1224 suword(&job->_aiocb_private.kernelinfo, 0);
1228 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1230 suword(&job->_aiocb_private.status, 0);
1235 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1236 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1237 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1241 * This method for requesting kevent-based notification won't
1242 * work on the alpha, since we're passing in a pointer
1243 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1244 * based method instead.
1246 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1247 user_opcode == LIO_WRITE)
1250 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1255 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1256 (kq_fp = fdp->fd_files[kev.ident].fp) == NULL ||
1257 (kq_fp->f_type != DTYPE_KQUEUE)) {
1261 kq = (struct kqueue *)kq_fp->f_data;
1262 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1263 kev.filter = EVFILT_AIO;
1264 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1265 kev.data = (intptr_t)aiocbe;
1266 /* XXX lwp kqueue_register takes a thread, but only uses its proc */
1267 error = kqueue_register(kq, &kev, FIRST_LWP_IN_PROC(p)->lwp_thread);
1271 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1273 suword(&job->_aiocb_private.error, error);
1278 suword(&job->_aiocb_private.error, EINPROGRESS);
1279 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1280 aiocbe->userproc = p;
1281 aiocbe->jobflags = 0;
1285 if (fp->f_type == DTYPE_SOCKET) {
1287 * Alternate queueing for socket ops: Reach down into the
1288 * descriptor to get the socket data. Then check to see if the
1289 * socket is ready to be read or written (based on the requested
1292 * If it is not ready for io, then queue the aiocbe on the
1293 * socket, and set the flags so we get a call when sbnotify()
1296 so = (struct socket *)fp->f_data;
1298 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1299 LIO_WRITE) && (!sowriteable(so)))) {
1300 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1301 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1302 if (opcode == LIO_READ)
1303 so->so_rcv.sb_flags |= SB_AIO;
1305 so->so_snd.sb_flags |= SB_AIO;
1306 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1307 ki->kaio_queue_count++;
1316 if ((error = aio_qphysio(p, aiocbe)) == 0)
1319 suword(&job->_aiocb_private.status, 0);
1320 aiocbe->uaiocb._aiocb_private.error = error;
1321 suword(&job->_aiocb_private.error, error);
1325 /* No buffer for daemon I/O. */
1328 ki->kaio_queue_count++;
1330 lj->lioj_queue_count++;
1332 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1333 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1335 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1341 * If we don't have a free AIO process, and we are below our quota, then
1342 * start one. Otherwise, depend on the subsequent I/O completions to
1343 * pick-up this job. If we don't successfully create the new process
1344 * (thread) due to resource issues, we return an error for now (EAGAIN),
1345 * which is likely not the correct thing to do.
1349 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1350 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1351 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1352 aiop->aioprocflags &= ~AIOP_FREE;
1353 wakeup(aiop->aioproc);
1354 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1355 ((ki->kaio_active_count + num_aio_resv_start) <
1356 ki->kaio_maxactive_count)) {
1357 num_aio_resv_start++;
1358 if ((error = aio_newproc()) == 0) {
1359 num_aio_resv_start--;
1362 num_aio_resv_start--;
1370 * This routine queues an AIO request, checking for quotas.
1373 aio_aqueue(struct aiocb *job, int type)
1375 struct proc *p = curproc;
1376 struct kaioinfo *ki;
1378 if (p->p_aioinfo == NULL)
1379 aio_init_aioinfo(p);
1381 if (num_queue_count >= max_queue_count)
1385 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1388 return _aio_aqueue(job, NULL, type);
1390 #endif /* VFS_AIO */
1393 * Support the aio_return system call, as a side-effect, kernel resources are
1397 sys_aio_return(struct aio_return_args *uap)
1402 struct proc *p = curproc;
1403 struct lwp *lp = curthread->td_lwp;
1405 struct aiocblist *cb, *ncb;
1407 struct kaioinfo *ki;
1415 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1416 if (jobref == -1 || jobref == 0)
1419 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1420 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1422 if (ujob == cb->uuaiocb) {
1423 uap->sysmsg_result =
1424 cb->uaiocb._aiocb_private.status;
1426 uap->sysmsg_result = EFAULT;
1427 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1428 lp->lwp_ru.ru_oublock += cb->outputcharge;
1429 cb->outputcharge = 0;
1430 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1431 lp->lwp_ru.ru_inblock += cb->inputcharge;
1432 cb->inputcharge = 0;
1439 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1440 ncb = TAILQ_NEXT(cb, plist);
1441 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1444 if (ujob == cb->uuaiocb) {
1445 uap->sysmsg_result =
1446 cb->uaiocb._aiocb_private.status;
1448 uap->sysmsg_result = EFAULT;
1456 #endif /* VFS_AIO */
1460 * Allow a process to wakeup when any of the I/O requests are completed.
1463 sys_aio_suspend(struct aio_suspend_args *uap)
1468 struct proc *p = curproc;
1471 struct aiocb *const *cbptr, *cbp;
1472 struct kaioinfo *ki;
1473 struct aiocblist *cb;
1478 struct aiocb **ujoblist;
1480 if (uap->nent > AIO_LISTIO_MAX)
1485 /* Get timespec struct. */
1486 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1489 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1492 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1493 if (itimerfix(&atv))
1495 timo = tvtohz_high(&atv);
1503 ijoblist = zalloc(aiol_zone);
1504 ujoblist = zalloc(aiol_zone);
1505 cbptr = uap->aiocbp;
1507 for (i = 0; i < uap->nent; i++) {
1508 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1511 ujoblist[njoblist] = cbp;
1512 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1516 if (njoblist == 0) {
1517 zfree(aiol_zone, ijoblist);
1518 zfree(aiol_zone, ujoblist);
1524 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1525 for (i = 0; i < njoblist; i++) {
1527 cb->uaiocb._aiocb_private.kernelinfo) ==
1529 if (ujoblist[i] != cb->uuaiocb)
1531 zfree(aiol_zone, ijoblist);
1532 zfree(aiol_zone, ujoblist);
1539 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1540 TAILQ_NEXT(cb, plist)) {
1541 for (i = 0; i < njoblist; i++) {
1543 cb->uaiocb._aiocb_private.kernelinfo) ==
1546 if (ujoblist[i] != cb->uuaiocb)
1548 zfree(aiol_zone, ijoblist);
1549 zfree(aiol_zone, ujoblist);
1555 ki->kaio_flags |= KAIO_WAKEUP;
1556 error = tsleep(p, PCATCH, "aiospn", timo);
1559 if (error == ERESTART || error == EINTR) {
1560 zfree(aiol_zone, ijoblist);
1561 zfree(aiol_zone, ujoblist);
1563 } else if (error == EWOULDBLOCK) {
1564 zfree(aiol_zone, ijoblist);
1565 zfree(aiol_zone, ujoblist);
1572 #endif /* VFS_AIO */
1576 * aio_cancel cancels any non-physio aio operations not currently in
1580 sys_aio_cancel(struct aio_cancel_args *uap)
1585 struct proc *p = curproc;
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_files[uap->fd].fp) == NULL)
1602 if (fp->f_type == DTYPE_VNODE) {
1603 vp = (struct vnode *)fp->f_data;
1605 if (vn_isdisk(vp,&error)) {
1606 uap->sysmsg_result = 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 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1640 if ((cancelled) && (uap->aiocbp)) {
1641 uap->sysmsg_result = 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 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1679 uap->sysmsg_result = AIO_NOTCANCELED;
1683 uap->sysmsg_result = AIO_CANCELED;
1686 uap->sysmsg_result = 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 sys_aio_error(struct aio_error_args *uap)
1703 struct proc *p = curproc;
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 uap->sysmsg_result = 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 uap->sysmsg_result = EINPROGRESS;
1736 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1738 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1740 uap->sysmsg_result = EINPROGRESS;
1748 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1750 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1752 uap->sysmsg_result = 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 uap->sysmsg_result = 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 sys_aio_read(struct aio_read_args *uap)
1788 return aio_aqueue(uap->aiocbp, LIO_READ);
1789 #endif /* VFS_AIO */
1792 /* syscall - asynchronous write to a file (REALTIME) */
1794 sys_aio_write(struct aio_write_args *uap)
1799 return aio_aqueue(uap->aiocbp, LIO_WRITE);
1800 #endif /* VFS_AIO */
1803 /* syscall - XXX undocumented */
1805 sys_lio_listio(struct lio_listio_args *uap)
1810 struct proc *p = curproc;
1811 struct lwp *lp = curthread->td_lwp;
1812 int nent, nentqueued;
1813 struct aiocb *iocb, * const *cbptr;
1814 struct aiocblist *cb;
1815 struct kaioinfo *ki;
1816 struct aio_liojob *lj;
1817 int error, runningcode;
1821 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1825 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)
1835 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1838 lj = zalloc(aiolio_zone);
1843 lj->lioj_buffer_count = 0;
1844 lj->lioj_buffer_finished_count = 0;
1845 lj->lioj_queue_count = 0;
1846 lj->lioj_queue_finished_count = 0;
1852 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1853 error = copyin(uap->sig, &lj->lioj_signal,
1854 sizeof(lj->lioj_signal));
1856 zfree(aiolio_zone, lj);
1859 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1860 zfree(aiolio_zone, lj);
1863 lj->lioj_flags |= LIOJ_SIGNAL;
1864 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1866 lj->lioj_flags &= ~LIOJ_SIGNAL;
1868 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1870 * Get pointers to the list of I/O requests.
1874 cbptr = uap->acb_list;
1875 for (i = 0; i < uap->nent; i++) {
1876 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1877 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1878 error = _aio_aqueue(iocb, lj, 0);
1887 * If we haven't queued any, then just return error.
1889 if (nentqueued == 0)
1893 * Calculate the appropriate error return.
1899 if (uap->mode == LIO_WAIT) {
1900 int command, found, jobref;
1904 for (i = 0; i < uap->nent; i++) {
1906 * Fetch address of the control buf pointer in
1909 iocb = (struct aiocb *)
1910 (intptr_t)fuword(&cbptr[i]);
1911 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1916 * Fetch the associated command from user space.
1918 command = fuword(&iocb->aio_lio_opcode);
1919 if (command == LIO_NOP) {
1924 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1926 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1927 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1929 if (cb->uaiocb.aio_lio_opcode
1931 lp->lwp_ru.ru_oublock +=
1933 cb->outputcharge = 0;
1934 } else if (cb->uaiocb.aio_lio_opcode
1936 lp->lwp_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, 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 ksignal(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 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2000 * Interrupt handler for physio, performs the necessary process wakeups, and
2004 aio_physwakeup(struct bio *bio)
2006 struct buf *bp = bio->bio_buf;
2007 struct aiocblist *aiocbe;
2009 struct kaioinfo *ki;
2010 struct aio_liojob *lj;
2012 aiocbe = bio->bio_caller_info2.ptr;
2015 p = bio->bio_caller_info1.ptr;
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 callout_reset(&aiocbe->timeout, 0,
2042 process_signal, aiocbe);
2049 ki->kaio_buffer_finished_count++;
2050 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2051 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2052 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2054 KNOTE(&aiocbe->klist, 0);
2055 /* Do the wakeup. */
2056 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2057 ki->kaio_flags &= ~KAIO_WAKEUP;
2062 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2063 callout_reset(&aiocbe->timeout, 0,
2064 process_signal, aiocbe);
2067 bp->b_cmd = BUF_CMD_DONE;
2070 #endif /* VFS_AIO */
2072 /* syscall - wait for the next completion of an aio request */
2074 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2079 struct proc *p = curproc;
2080 struct lwp *lp = curthread->td_lwp;
2083 struct kaioinfo *ki;
2084 struct aiocblist *cb = NULL;
2087 suword(uap->aiocbp, (int)NULL);
2091 /* Get timespec struct. */
2092 error = copyin(uap->timeout, &ts, sizeof(ts));
2096 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2099 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2100 if (itimerfix(&atv))
2102 timo = tvtohz_high(&atv);
2110 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2111 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2112 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2113 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2114 lp->lwp_ru.ru_oublock +=
2116 cb->outputcharge = 0;
2117 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2118 lp->lwp_ru.ru_inblock += cb->inputcharge;
2119 cb->inputcharge = 0;
2122 return cb->uaiocb._aiocb_private.error;
2126 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2128 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2129 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2131 return cb->uaiocb._aiocb_private.error;
2134 ki->kaio_flags |= KAIO_WAKEUP;
2135 error = tsleep(p, PCATCH, "aiowc", timo);
2138 if (error == ERESTART)
2142 else if (error == EINTR)
2144 else if (error == EWOULDBLOCK)
2147 #endif /* VFS_AIO */
2152 filt_aioattach(struct knote *kn)
2158 struct filterops aio_filtops =
2159 { 0, filt_aioattach, NULL, NULL };
2162 /* kqueue attach function */
2164 filt_aioattach(struct knote *kn)
2166 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2169 * The aiocbe pointer must be validated before using it, so
2170 * registration is restricted to the kernel; the user cannot
2173 if ((kn->kn_flags & EV_FLAG1) == 0)
2175 kn->kn_flags &= ~EV_FLAG1;
2177 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2182 /* kqueue detach function */
2184 filt_aiodetach(struct knote *kn)
2186 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2188 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2191 /* kqueue filter function */
2194 filt_aio(struct knote *kn, long hint)
2196 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2198 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2199 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2200 aiocbe->jobstate != JOBST_JOBBFINISHED)
2202 kn->kn_flags |= EV_EOF;
2206 struct filterops aio_filtops =
2207 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2208 #endif /* VFS_AIO */