2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
7 * 1. Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * 2. John S. Dyson's name may not be used to endorse or promote products
10 * derived from this software without specific prior written permission.
12 * DISCLAIMER: This code isn't warranted to do anything useful. Anything
13 * bad that happens because of using this software isn't the responsibility
14 * of the author. This software is distributed AS-IS.
16 * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.42 2007/07/20 17:21:52 dillon Exp $
21 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
24 #include <sys/param.h>
25 #include <sys/systm.h>
27 #include <sys/sysproto.h>
28 #include <sys/filedesc.h>
29 #include <sys/kernel.h>
30 #include <sys/fcntl.h>
33 #include <sys/unistd.h>
35 #include <sys/resourcevar.h>
36 #include <sys/signalvar.h>
37 #include <sys/protosw.h>
38 #include <sys/socketvar.h>
39 #include <sys/sysctl.h>
40 #include <sys/vnode.h>
42 #include <sys/event.h>
45 #include <vm/vm_extern.h>
47 #include <vm/vm_map.h>
48 #include <vm/vm_zone.h>
51 #include <sys/file2.h>
53 #include <sys/sysref2.h>
54 #include <sys/thread2.h>
55 #include <sys/mplock2.h>
57 #include <machine/limits.h>
58 #include "opt_vfs_aio.h"
63 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
68 #define JOBST_NULL 0x0
69 #define JOBST_JOBQGLOBAL 0x2
70 #define JOBST_JOBRUNNING 0x3
71 #define JOBST_JOBFINISHED 0x4
72 #define JOBST_JOBQBUF 0x5
73 #define JOBST_JOBBFINISHED 0x6
75 #ifndef MAX_AIO_PER_PROC
76 #define MAX_AIO_PER_PROC 32
79 #ifndef MAX_AIO_QUEUE_PER_PROC
80 #define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
84 #define MAX_AIO_PROCS 32
88 #define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
91 #ifndef TARGET_AIO_PROCS
92 #define TARGET_AIO_PROCS 4
96 #define MAX_BUF_AIO 16
99 #ifndef AIOD_TIMEOUT_DEFAULT
100 #define AIOD_TIMEOUT_DEFAULT (10 * hz)
103 #ifndef AIOD_LIFETIME_DEFAULT
104 #define AIOD_LIFETIME_DEFAULT (30 * hz)
107 SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
109 static int max_aio_procs = MAX_AIO_PROCS;
110 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
111 CTLFLAG_RW, &max_aio_procs, 0,
112 "Maximum number of kernel threads to use for handling async IO");
114 static int num_aio_procs = 0;
115 SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
116 CTLFLAG_RD, &num_aio_procs, 0,
117 "Number of presently active kernel threads for async IO");
120 * The code will adjust the actual number of AIO processes towards this
121 * number when it gets a chance.
123 static int target_aio_procs = TARGET_AIO_PROCS;
124 SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
125 0, "Preferred number of ready kernel threads for async IO");
127 static int max_queue_count = MAX_AIO_QUEUE;
128 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
129 "Maximum number of aio requests to queue, globally");
131 static int num_queue_count = 0;
132 SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
133 "Number of queued aio requests");
135 static int num_buf_aio = 0;
136 SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
137 "Number of aio requests presently handled by the buf subsystem");
139 /* Number of async I/O thread in the process of being started */
140 /* XXX This should be local to _aio_aqueue() */
141 static int num_aio_resv_start = 0;
143 static int aiod_timeout;
144 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
145 "Timeout value for synchronous aio operations");
147 static int aiod_lifetime;
148 SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
149 "Maximum lifetime for idle aiod");
151 static int max_aio_per_proc = MAX_AIO_PER_PROC;
152 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
153 0, "Maximum active aio requests per process (stored in the process)");
155 static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
156 SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
157 &max_aio_queue_per_proc, 0,
158 "Maximum queued aio requests per process (stored in the process)");
160 static int max_buf_aio = MAX_BUF_AIO;
161 SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
162 "Maximum buf aio requests per process (stored in the process)");
167 #define AIOP_FREE 0x1 /* proc on free queue */
168 #define AIOP_SCHED 0x2 /* proc explicitly scheduled */
171 int aioprocflags; /* AIO proc flags */
172 TAILQ_ENTRY(aioproclist) list; /* List of processes */
173 struct proc *aioproc; /* The AIO thread */
177 * data-structure for lio signal management
181 int lioj_buffer_count;
182 int lioj_buffer_finished_count;
183 int lioj_queue_count;
184 int lioj_queue_finished_count;
185 struct sigevent lioj_signal; /* signal on all I/O done */
186 TAILQ_ENTRY(aio_liojob) lioj_list;
187 struct kaioinfo *lioj_ki;
189 #define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
190 #define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
193 * per process aio data structure
196 int kaio_flags; /* per process kaio flags */
197 int kaio_maxactive_count; /* maximum number of AIOs */
198 int kaio_active_count; /* number of currently used AIOs */
199 int kaio_qallowed_count; /* maxiumu size of AIO queue */
200 int kaio_queue_count; /* size of AIO queue */
201 int kaio_ballowed_count; /* maximum number of buffers */
202 int kaio_queue_finished_count; /* number of daemon jobs finished */
203 int kaio_buffer_count; /* number of physio buffers */
204 int kaio_buffer_finished_count; /* count of I/O done */
205 struct proc *kaio_p; /* process that uses this kaio block */
206 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
207 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
208 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
209 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
210 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
211 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
214 #define KAIO_RUNDOWN 0x1 /* process is being run down */
215 #define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
217 static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
218 static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
219 static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
220 static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
222 static void aio_init_aioinfo(struct proc *p);
223 static void aio_onceonly(void *);
224 static int aio_free_entry(struct aiocblist *aiocbe);
225 static void aio_process(struct aiocblist *aiocbe);
226 static int aio_newproc(void);
227 static int aio_aqueue(struct aiocb *job, int type);
228 static void aio_physwakeup(struct bio *bio);
229 static int aio_fphysio(struct aiocblist *aiocbe);
230 static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
231 static void aio_daemon(void *uproc, struct trapframe *frame);
232 static void process_signal(void *aioj);
234 SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
238 * kaio Per process async io info
239 * aiop async io thread data
240 * aiocb async io jobs
241 * aiol list io job pointer - internal to aio_suspend XXX
242 * aiolio list io jobs
244 static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
247 * Startup initialization
250 aio_onceonly(void *na)
252 TAILQ_INIT(&aio_freeproc);
253 TAILQ_INIT(&aio_activeproc);
254 TAILQ_INIT(&aio_jobs);
255 TAILQ_INIT(&aio_bufjobs);
256 TAILQ_INIT(&aio_freejobs);
257 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
258 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
259 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
260 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
261 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
262 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
263 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
268 * Init the per-process aioinfo structure. The aioinfo limits are set
269 * per-process for user limit (resource) management.
272 aio_init_aioinfo(struct proc *p)
275 if (p->p_aioinfo == NULL) {
276 ki = zalloc(kaio_zone);
279 ki->kaio_maxactive_count = max_aio_per_proc;
280 ki->kaio_active_count = 0;
281 ki->kaio_qallowed_count = max_aio_queue_per_proc;
282 ki->kaio_queue_count = 0;
283 ki->kaio_ballowed_count = max_buf_aio;
284 ki->kaio_buffer_count = 0;
285 ki->kaio_buffer_finished_count = 0;
287 TAILQ_INIT(&ki->kaio_jobdone);
288 TAILQ_INIT(&ki->kaio_jobqueue);
289 TAILQ_INIT(&ki->kaio_bufdone);
290 TAILQ_INIT(&ki->kaio_bufqueue);
291 TAILQ_INIT(&ki->kaio_liojoblist);
292 TAILQ_INIT(&ki->kaio_sockqueue);
295 while (num_aio_procs < target_aio_procs)
300 * Free a job entry. Wait for completion if it is currently active, but don't
301 * delay forever. If we delay, we return a flag that says that we have to
302 * restart the queue scan.
305 aio_free_entry(struct aiocblist *aiocbe)
308 struct aio_liojob *lj;
312 if (aiocbe->jobstate == JOBST_NULL)
313 panic("aio_free_entry: freeing already free job");
315 p = aiocbe->userproc;
319 panic("aio_free_entry: missing p->p_aioinfo");
321 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
322 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
323 tsleep(aiocbe, 0, "jobwai", 0);
325 if (aiocbe->bp == NULL) {
326 if (ki->kaio_queue_count <= 0)
327 panic("aio_free_entry: process queue size <= 0");
328 if (num_queue_count <= 0)
329 panic("aio_free_entry: system wide queue size <= 0");
332 lj->lioj_queue_count--;
333 if (aiocbe->jobflags & AIOCBLIST_DONE)
334 lj->lioj_queue_finished_count--;
336 ki->kaio_queue_count--;
337 if (aiocbe->jobflags & AIOCBLIST_DONE)
338 ki->kaio_queue_finished_count--;
342 lj->lioj_buffer_count--;
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 lj->lioj_buffer_finished_count--;
346 if (aiocbe->jobflags & AIOCBLIST_DONE)
347 ki->kaio_buffer_finished_count--;
348 ki->kaio_buffer_count--;
352 /* aiocbe is going away, we need to destroy any knotes */
353 /* XXX lwp knote wants a thread, but only cares about the process */
354 knote_empty(&aiocbe->klist);
356 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
357 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
358 ki->kaio_flags &= ~KAIO_WAKEUP;
362 if (aiocbe->jobstate == JOBST_JOBQBUF) {
363 if ((error = aio_fphysio(aiocbe)) != 0)
365 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
366 panic("aio_free_entry: invalid physio finish-up state");
368 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
370 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
372 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
373 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
375 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
376 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
377 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
379 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
382 vunmapbuf(aiocbe->bp);
383 relpbuf(aiocbe->bp, NULL);
387 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
388 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
389 zfree(aiolio_zone, lj);
391 aiocbe->jobstate = JOBST_NULL;
392 callout_stop(&aiocbe->timeout);
393 fdrop(aiocbe->fd_file);
394 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
400 * Rundown the jobs for a given process.
403 aio_proc_rundown(struct proc *p)
409 struct aio_liojob *lj, *ljn;
410 struct aiocblist *aiocbe, *aiocbn;
418 ki->kaio_flags |= LIOJ_SIGNAL_POSTED;
419 while ((ki->kaio_active_count > 0) || (ki->kaio_buffer_count >
420 ki->kaio_buffer_finished_count)) {
421 ki->kaio_flags |= KAIO_RUNDOWN;
422 if (tsleep(p, 0, "kaiowt", aiod_timeout))
427 * Move any aio ops that are waiting on socket I/O to the normal job
428 * queues so they are cleaned up with any others.
431 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
433 aiocbn = TAILQ_NEXT(aiocbe, plist);
434 fp = aiocbe->fd_file;
436 so = (struct socket *)fp->f_data;
437 TAILQ_REMOVE(&so->so_aiojobq, aiocbe, list);
438 if (TAILQ_EMPTY(&so->so_aiojobq)) {
439 atomic_clear_int(&so->so_snd.ssb_flags,
441 atomic_clear_int(&so->so_rcv.ssb_flags,
445 TAILQ_REMOVE(&ki->kaio_sockqueue, aiocbe, plist);
446 TAILQ_INSERT_HEAD(&aio_jobs, aiocbe, list);
447 TAILQ_INSERT_HEAD(&ki->kaio_jobqueue, aiocbe, plist);
452 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
453 aiocbn = TAILQ_NEXT(aiocbe, plist);
454 if (aio_free_entry(aiocbe))
459 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
461 aiocbn = TAILQ_NEXT(aiocbe, plist);
462 if (aio_free_entry(aiocbe))
468 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
469 ki->kaio_flags |= KAIO_WAKEUP;
470 tsleep(p, 0, "aioprn", 0);
478 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
479 aiocbn = TAILQ_NEXT(aiocbe, plist);
480 if (aio_free_entry(aiocbe)) {
488 * If we've slept, jobs might have moved from one queue to another.
489 * Retry rundown if we didn't manage to empty the queues.
491 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
492 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
493 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
494 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
497 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
498 ljn = TAILQ_NEXT(lj, lioj_list);
499 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
501 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
502 zfree(aiolio_zone, lj);
505 kprintf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
506 "QF:%d\n", lj->lioj_buffer_count,
507 lj->lioj_buffer_finished_count,
508 lj->lioj_queue_count,
509 lj->lioj_queue_finished_count);
514 zfree(kaio_zone, ki);
521 * Select a job to run (called by an AIO daemon).
523 static struct aiocblist *
524 aio_selectjob(struct aioproclist *aiop)
526 struct aiocblist *aiocbe;
531 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
532 TAILQ_NEXT(aiocbe, list)) {
533 userp = aiocbe->userproc;
534 ki = userp->p_aioinfo;
536 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
537 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
548 * The AIO processing activity. This is the code that does the I/O request for
549 * the non-physio version of the operations. The normal vn operations are used,
550 * and this code should work in all instances for every type of file, including
551 * pipes, sockets, fifos, and regular files.
554 aio_process(struct aiocblist *aiocbe)
563 int oublock_st, oublock_end;
564 int inblock_st, inblock_end;
567 cb = &aiocbe->uaiocb;
568 fp = aiocbe->fd_file;
570 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
571 aiov.iov_len = cb->aio_nbytes;
573 auio.uio_iov = &aiov;
575 auio.uio_offset = cb->aio_offset;
576 auio.uio_resid = cb->aio_nbytes;
577 cnt = cb->aio_nbytes;
578 auio.uio_segflg = UIO_USERSPACE;
581 inblock_st = mytd->td_lwp->lwp_ru.ru_inblock;
582 oublock_st = mytd->td_lwp->lwp_ru.ru_oublock;
584 * _aio_aqueue() acquires a reference to the file that is
585 * released in aio_free_entry().
587 if (cb->aio_lio_opcode == LIO_READ) {
588 auio.uio_rw = UIO_READ;
589 error = fo_read(fp, &auio, fp->f_cred, O_FOFFSET);
591 auio.uio_rw = UIO_WRITE;
592 error = fo_write(fp, &auio, fp->f_cred, O_FOFFSET);
594 inblock_end = mytd->td_lwp->lwp_ru.ru_inblock;
595 oublock_end = mytd->td_lwp->lwp_ru.ru_oublock;
597 aiocbe->inputcharge = inblock_end - inblock_st;
598 aiocbe->outputcharge = oublock_end - oublock_st;
600 if ((error) && (auio.uio_resid != cnt)) {
601 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
603 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
604 ksignal(aiocbe->userproc, SIGPIPE);
607 cnt -= auio.uio_resid;
608 cb->_aiocb_private.error = error;
609 cb->_aiocb_private.status = cnt;
613 * The AIO daemon, most of the actual work is done in aio_process,
614 * but the setup (and address space mgmt) is done in this routine.
617 aio_daemon(void *uproc, struct trapframe *frame)
619 struct aio_liojob *lj;
621 struct aiocblist *aiocbe;
622 struct aioproclist *aiop;
624 struct proc *mycp, *userp;
625 struct vmspace *curvm;
630 * mplock not held on entry but we aren't mpsafe yet.
634 mylwp = curthread->td_lwp;
635 mycp = mylwp->lwp_proc;
637 if (mycp->p_textvp) {
638 vrele(mycp->p_textvp);
639 mycp->p_textvp = NULL;
643 * Allocate and ready the aio control info. There is one aiop structure
646 aiop = zalloc(aiop_zone);
647 aiop->aioproc = mycp;
648 aiop->aioprocflags |= AIOP_FREE;
653 * Place thread (lightweight process) onto the AIO free thread list.
655 if (TAILQ_EMPTY(&aio_freeproc))
656 wakeup(&aio_freeproc);
657 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
661 /* Make up a name for the daemon. */
662 strcpy(mycp->p_comm, "aiod");
665 * Get rid of our current filedescriptors. AIOD's don't need any
666 * filedescriptors, except as temporarily inherited from the client.
667 * Credentials are also cloned, and made equivalent to "root".
670 cr = cratom(&mycp->p_ucred);
672 uireplace(&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.)
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 (curvm != userp->p_vmspace) {
715 pmap_setlwpvm(mylwp, userp->p_vmspace);
717 sysref_put(&curvm->vm_sysref);
718 curvm = userp->p_vmspace;
719 sysref_get(&curvm->vm_sysref);
722 ki = userp->p_aioinfo;
725 /* Account for currently active jobs. */
726 ki->kaio_active_count++;
728 /* Do the I/O function. */
731 /* Decrement the active job count. */
732 ki->kaio_active_count--;
735 * Increment the completion count for wakeup/signal
738 aiocbe->jobflags |= AIOCBLIST_DONE;
739 ki->kaio_queue_finished_count++;
741 lj->lioj_queue_finished_count++;
742 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
743 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
744 ki->kaio_flags &= ~KAIO_WAKEUP;
749 if (lj && (lj->lioj_flags &
750 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
751 if ((lj->lioj_queue_finished_count ==
752 lj->lioj_queue_count) &&
753 (lj->lioj_buffer_finished_count ==
754 lj->lioj_buffer_count)) {
756 lj->lioj_signal.sigev_signo);
763 aiocbe->jobstate = JOBST_JOBFINISHED;
766 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
767 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
769 KNOTE(&aiocbe->klist, 0);
771 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
773 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
776 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
777 ksignal(userp, cb->aio_sigevent.sigev_signo);
782 * Disconnect from user address space.
785 /* swap our original address space back in */
786 pmap_setlwpvm(mylwp, mycp->p_vmspace);
787 sysref_put(&curvm->vm_sysref);
792 * If we are the first to be put onto the free queue, wakeup
793 * anyone waiting for a daemon.
796 TAILQ_REMOVE(&aio_activeproc, aiop, list);
797 if (TAILQ_EMPTY(&aio_freeproc))
798 wakeup(&aio_freeproc);
799 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
800 aiop->aioprocflags |= AIOP_FREE;
804 * If daemon is inactive for a long time, allow it to exit,
805 * thereby freeing resources.
807 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
808 0, "aiordy", aiod_lifetime)) {
810 if (TAILQ_EMPTY(&aio_jobs)) {
811 if ((aiop->aioprocflags & AIOP_FREE) &&
812 (num_aio_procs > target_aio_procs)) {
813 TAILQ_REMOVE(&aio_freeproc, aiop, list);
815 zfree(aiop_zone, aiop);
818 if (mycp->p_vmspace->vm_sysref.refcnt <= 1) {
819 kprintf("AIOD: bad vm refcnt for"
820 " exiting daemon: %d\n",
821 mycp->p_vmspace->vm_sysref.refcnt);
833 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
834 * AIO daemon modifies its environment itself.
840 struct lwp *lp, *nlp;
844 error = fork1(lp, RFPROC|RFMEM|RFNOWAIT, &np);
847 nlp = ONLY_LWP_IN_PROC(np);
848 cpu_set_fork_handler(nlp, aio_daemon, curproc);
849 start_forked_proc(lp, np);
852 * Wait until daemon is started, but continue on just in case to
853 * handle error conditions.
855 error = tsleep(np, 0, "aiosta", aiod_timeout);
862 * Try the high-performance, low-overhead physio method for eligible
863 * VCHR devices. This method doesn't use an aio helper thread, and
864 * thus has very low overhead.
866 * Assumes that the caller, _aio_aqueue(), has incremented the file
867 * structure's reference count, preventing its deallocation for the
868 * duration of this call.
871 aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
879 struct aio_liojob *lj;
882 cb = &aiocbe->uaiocb;
883 fp = aiocbe->fd_file;
885 if (fp->f_type != DTYPE_VNODE)
888 vp = (struct vnode *)fp->f_data;
891 * If its not a disk, we don't want to return a positive error.
892 * It causes the aio code to not fall through to try the thread
893 * way when you're talking to a regular file.
895 if (!vn_isdisk(vp, &error)) {
896 if (error == ENOTBLK)
902 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
906 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
910 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
913 ki->kaio_buffer_count++;
917 lj->lioj_buffer_count++;
919 /* Create and build a buffer header for a transfer. */
920 bp = getpbuf_kva(NULL);
924 * Get a copy of the kva from the physical buffer.
926 bp->b_bio1.bio_caller_info1.ptr = p;
929 bp->b_cmd = (cb->aio_lio_opcode == LIO_WRITE) ?
930 BUF_CMD_WRITE : BUF_CMD_READ;
931 bp->b_bio1.bio_done = aio_physwakeup;
932 bp->b_bio1.bio_flags |= BIO_SYNC;
933 bp->b_bio1.bio_offset = cb->aio_offset;
935 /* Bring buffer into kernel space. */
936 if (vmapbuf(bp, __DEVOLATILE(char *, cb->aio_buf), cb->aio_nbytes) < 0) {
944 bp->b_bio1.bio_caller_info2.ptr = aiocbe;
945 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
946 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
947 aiocbe->jobstate = JOBST_JOBQBUF;
948 cb->_aiocb_private.status = cb->aio_nbytes;
955 * Perform the transfer. vn_strategy must be used even though we
956 * know we have a device in order to deal with requests which exceed
957 * device DMA limitations.
959 vn_strategy(vp, &bp->b_bio1);
966 * If we had an error invoking the request, or an error in processing
967 * the request before we have returned, we process it as an error in
968 * transfer. Note that such an I/O error is not indicated immediately,
969 * but is returned using the aio_error mechanism. In this case,
970 * aio_suspend will return immediately.
972 if (bp->b_error || (bp->b_flags & B_ERROR)) {
973 struct aiocb *job = aiocbe->uuaiocb;
975 aiocbe->uaiocb._aiocb_private.status = 0;
976 suword(&job->_aiocb_private.status, 0);
977 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
978 suword(&job->_aiocb_private.error, bp->b_error);
980 ki->kaio_buffer_finished_count++;
982 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
983 aiocbe->jobstate = JOBST_JOBBFINISHED;
984 aiocbe->jobflags |= AIOCBLIST_DONE;
985 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
986 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
987 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
994 KNOTE(&aiocbe->klist, 0);
998 ki->kaio_buffer_count--;
1000 lj->lioj_buffer_count--;
1007 * This waits/tests physio completion.
1010 aio_fphysio(struct aiocblist *iocb)
1017 error = biowait_timeout(&bp->b_bio1, "physstr", aiod_timeout);
1018 if (error == EWOULDBLOCK)
1021 /* Release mapping into kernel space. */
1027 /* Check for an error. */
1028 if (bp->b_flags & B_ERROR)
1029 error = bp->b_error;
1034 #endif /* VFS_AIO */
1037 * Wake up aio requests that may be serviceable now.
1040 aio_swake(struct socket *so, struct signalsockbuf *ssb)
1045 struct aiocblist *cb,*cbn;
1047 struct kaioinfo *ki = NULL;
1048 int opcode, wakecount = 0;
1049 struct aioproclist *aiop;
1051 if (ssb == &so->so_snd) {
1053 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AIO);
1056 atomic_clear_int(&so->so_rcv.ssb_flags, SSB_AIO);
1059 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1060 cbn = TAILQ_NEXT(cb, list);
1061 if (opcode == cb->uaiocb.aio_lio_opcode) {
1064 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1065 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1066 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1067 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1069 if (cb->jobstate != JOBST_JOBQGLOBAL)
1070 panic("invalid queue value");
1074 while (wakecount--) {
1075 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1076 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1077 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1078 aiop->aioprocflags &= ~AIOP_FREE;
1079 wakeup(aiop->aioproc);
1082 #endif /* VFS_AIO */
1087 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1088 * technique is done in this code.
1091 _aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
1093 struct proc *p = curproc;
1098 int opcode, user_opcode;
1099 struct aiocblist *aiocbe;
1100 struct aioproclist *aiop;
1101 struct kaioinfo *ki;
1107 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1108 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1110 aiocbe = zalloc (aiocb_zone);
1112 aiocbe->inputcharge = 0;
1113 aiocbe->outputcharge = 0;
1114 callout_init(&aiocbe->timeout);
1115 SLIST_INIT(&aiocbe->klist);
1117 suword(&job->_aiocb_private.status, -1);
1118 suword(&job->_aiocb_private.error, 0);
1119 suword(&job->_aiocb_private.kernelinfo, -1);
1121 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1123 suword(&job->_aiocb_private.error, error);
1124 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1127 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1128 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1129 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1133 /* Save userspace address of the job info. */
1134 aiocbe->uuaiocb = job;
1136 /* Get the opcode. */
1137 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1138 if (type != LIO_NOP)
1139 aiocbe->uaiocb.aio_lio_opcode = type;
1140 opcode = aiocbe->uaiocb.aio_lio_opcode;
1143 * Range check file descriptor.
1145 fflags = (opcode == LIO_WRITE) ? FWRITE : FREAD;
1146 fd = aiocbe->uaiocb.aio_fildes;
1147 fp = holdfp(p->p_fd, fd, fflags);
1149 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1151 suword(&job->_aiocb_private.error, EBADF);
1155 aiocbe->fd_file = fp;
1157 if (aiocbe->uaiocb.aio_offset == -1LL) {
1161 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1166 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1167 if (jobrefid == LONG_MAX)
1172 if (opcode == LIO_NOP) {
1174 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1176 suword(&job->_aiocb_private.error, 0);
1177 suword(&job->_aiocb_private.status, 0);
1178 suword(&job->_aiocb_private.kernelinfo, 0);
1182 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1184 suword(&job->_aiocb_private.status, 0);
1189 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1190 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1191 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sival_ptr;
1195 * This method for requesting kevent-based notification won't
1196 * work on the alpha, since we're passing in a pointer
1197 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1198 * based method instead.
1200 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1201 user_opcode == LIO_WRITE)
1204 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1209 kq_fp = holdfp(p->p_fd, (int)kev.ident, -1);
1210 if (kq_fp == NULL || kq_fp->f_type != DTYPE_KQUEUE) {
1218 kq = (struct kqueue *)kq_fp->f_data;
1219 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1220 kev.filter = EVFILT_AIO;
1221 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1222 kev.data = (intptr_t)aiocbe;
1223 error = kqueue_register(kq, &kev);
1228 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1230 suword(&job->_aiocb_private.error, error);
1235 suword(&job->_aiocb_private.error, EINPROGRESS);
1236 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1237 aiocbe->userproc = p;
1238 aiocbe->jobflags = 0;
1242 if (fp->f_type == DTYPE_SOCKET) {
1244 * Alternate queueing for socket ops: Reach down into the
1245 * descriptor to get the socket data. Then check to see if the
1246 * socket is ready to be read or written (based on the requested
1249 * If it is not ready for io, then queue the aiocbe on the
1250 * socket, and set the flags so we get a call when ssb_notify()
1253 so = (struct socket *)fp->f_data;
1255 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1256 LIO_WRITE) && (!sowriteable(so)))) {
1257 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1258 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1259 if (opcode == LIO_READ)
1260 atomic_set_int(&so->so_rcv.ssb_flags, SSB_AIO);
1262 atomic_set_int(&so->so_snd.ssb_flags, SSB_AIO);
1263 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1264 ki->kaio_queue_count++;
1273 if ((error = aio_qphysio(p, aiocbe)) == 0)
1276 suword(&job->_aiocb_private.status, 0);
1277 aiocbe->uaiocb._aiocb_private.error = error;
1278 suword(&job->_aiocb_private.error, error);
1282 /* No buffer for daemon I/O. */
1285 ki->kaio_queue_count++;
1287 lj->lioj_queue_count++;
1289 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1290 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1292 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1298 * If we don't have a free AIO process, and we are below our quota, then
1299 * start one. Otherwise, depend on the subsequent I/O completions to
1300 * pick-up this job. If we don't successfully create the new process
1301 * (thread) due to resource issues, we return an error for now (EAGAIN),
1302 * which is likely not the correct thing to do.
1306 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1307 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1308 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1309 aiop->aioprocflags &= ~AIOP_FREE;
1310 wakeup(aiop->aioproc);
1311 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1312 ((ki->kaio_active_count + num_aio_resv_start) <
1313 ki->kaio_maxactive_count)) {
1314 num_aio_resv_start++;
1315 if ((error = aio_newproc()) == 0) {
1316 num_aio_resv_start--;
1319 num_aio_resv_start--;
1327 * This routine queues an AIO request, checking for quotas.
1330 aio_aqueue(struct aiocb *job, int type)
1332 struct proc *p = curproc;
1333 struct kaioinfo *ki;
1335 if (p->p_aioinfo == NULL)
1336 aio_init_aioinfo(p);
1338 if (num_queue_count >= max_queue_count)
1342 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1345 return _aio_aqueue(job, NULL, type);
1347 #endif /* VFS_AIO */
1350 * Support the aio_return system call, as a side-effect, kernel resources are
1356 sys_aio_return(struct aio_return_args *uap)
1361 struct proc *p = curproc;
1362 struct lwp *lp = curthread->td_lwp;
1364 struct aiocblist *cb, *ncb;
1366 struct kaioinfo *ki;
1375 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1376 if (jobref == -1 || jobref == 0)
1380 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1381 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1383 if (ujob == cb->uuaiocb) {
1384 uap->sysmsg_result =
1385 cb->uaiocb._aiocb_private.status;
1387 uap->sysmsg_result = EFAULT;
1389 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1390 lp->lwp_ru.ru_oublock += cb->outputcharge;
1391 cb->outputcharge = 0;
1392 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1393 lp->lwp_ru.ru_inblock += cb->inputcharge;
1394 cb->inputcharge = 0;
1402 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1403 ncb = TAILQ_NEXT(cb, plist);
1404 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1407 if (ujob == cb->uuaiocb) {
1408 uap->sysmsg_result =
1409 cb->uaiocb._aiocb_private.status;
1411 uap->sysmsg_result = EFAULT;
1423 #endif /* VFS_AIO */
1427 * Allow a process to wakeup when any of the I/O requests are completed.
1432 sys_aio_suspend(struct aio_suspend_args *uap)
1437 struct proc *p = curproc;
1440 struct aiocb *const *cbptr, *cbp;
1441 struct kaioinfo *ki;
1442 struct aiocblist *cb;
1447 struct aiocb **ujoblist;
1449 if ((u_int)uap->nent > AIO_LISTIO_MAX)
1454 /* Get timespec struct. */
1455 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1458 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1461 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1462 if (itimerfix(&atv))
1464 timo = tvtohz_high(&atv);
1474 ijoblist = zalloc(aiol_zone);
1475 ujoblist = zalloc(aiol_zone);
1476 cbptr = uap->aiocbp;
1478 for (i = 0; i < uap->nent; i++) {
1479 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1482 ujoblist[njoblist] = cbp;
1483 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1487 if (njoblist == 0) {
1488 zfree(aiol_zone, ijoblist);
1489 zfree(aiol_zone, ujoblist);
1496 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1497 for (i = 0; i < njoblist; i++) {
1499 cb->uaiocb._aiocb_private.kernelinfo) ==
1501 if (ujoblist[i] != cb->uuaiocb)
1503 zfree(aiol_zone, ijoblist);
1504 zfree(aiol_zone, ujoblist);
1511 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1512 TAILQ_NEXT(cb, plist)) {
1513 for (i = 0; i < njoblist; i++) {
1515 cb->uaiocb._aiocb_private.kernelinfo) ==
1518 if (ujoblist[i] != cb->uuaiocb)
1520 zfree(aiol_zone, ijoblist);
1521 zfree(aiol_zone, ujoblist);
1527 ki->kaio_flags |= KAIO_WAKEUP;
1528 error = tsleep(p, PCATCH, "aiospn", timo);
1531 if (error == ERESTART || error == EINTR) {
1532 zfree(aiol_zone, ijoblist);
1533 zfree(aiol_zone, ujoblist);
1536 } else if (error == EWOULDBLOCK) {
1537 zfree(aiol_zone, ijoblist);
1538 zfree(aiol_zone, ujoblist);
1549 #endif /* VFS_AIO */
1553 * aio_cancel cancels any non-physio aio operations not currently in
1559 sys_aio_cancel(struct aio_cancel_args *uap)
1564 struct proc *p = curproc;
1565 struct kaioinfo *ki;
1566 struct aiocblist *cbe, *cbn;
1575 fp = holdfp(p->p_fd, uap->fd, -1);
1581 if (fp->f_type == DTYPE_VNODE) {
1582 vp = (struct vnode *)fp->f_data;
1584 if (vn_isdisk(vp,&error)) {
1585 uap->sysmsg_result = AIO_NOTCANCELED;
1589 } else if (fp->f_type == DTYPE_SOCKET) {
1590 so = (struct socket *)fp->f_data;
1594 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1595 cbn = TAILQ_NEXT(cbe, list);
1596 if ((uap->aiocbp == NULL) ||
1597 (uap->aiocbp == cbe->uuaiocb) ) {
1600 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1601 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1602 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1603 if (ki->kaio_flags & KAIO_WAKEUP) {
1606 cbe->jobstate = JOBST_JOBFINISHED;
1607 cbe->uaiocb._aiocb_private.status=-1;
1608 cbe->uaiocb._aiocb_private.error=ECANCELED;
1610 /* XXX cancelled, knote? */
1611 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1613 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1620 if ((cancelled) && (uap->aiocbp)) {
1621 uap->sysmsg_result = AIO_CANCELED;
1631 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1632 cbn = TAILQ_NEXT(cbe, plist);
1634 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1635 ((uap->aiocbp == NULL ) ||
1636 (uap->aiocbp == cbe->uuaiocb))) {
1638 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1639 TAILQ_REMOVE(&aio_jobs, cbe, list);
1640 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1641 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1644 ki->kaio_queue_finished_count++;
1645 cbe->jobstate = JOBST_JOBFINISHED;
1646 cbe->uaiocb._aiocb_private.status = -1;
1647 cbe->uaiocb._aiocb_private.error = ECANCELED;
1648 /* XXX cancelled, knote? */
1649 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1651 ksignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1660 uap->sysmsg_result = AIO_NOTCANCELED;
1662 uap->sysmsg_result = AIO_CANCELED;
1664 uap->sysmsg_result = AIO_ALLDONE;
1670 #endif /* VFS_AIO */
1674 * aio_error is implemented in the kernel level for compatibility purposes only.
1675 * For a user mode async implementation, it would be best to do it in a userland
1681 sys_aio_error(struct aio_error_args *uap)
1686 struct proc *p = curproc;
1687 struct aiocblist *cb;
1688 struct kaioinfo *ki;
1696 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1697 if ((jobref == -1) || (jobref == 0))
1703 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1704 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1706 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1713 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1715 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1717 uap->sysmsg_result = EINPROGRESS;
1723 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1725 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1727 uap->sysmsg_result = EINPROGRESS;
1735 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1737 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1739 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
1745 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1747 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1749 uap->sysmsg_result = EINPROGRESS;
1759 #endif /* VFS_AIO */
1763 * syscall - asynchronous read from a file (REALTIME)
1768 sys_aio_read(struct aio_read_args *uap)
1776 error = aio_aqueue(uap->aiocbp, LIO_READ);
1779 #endif /* VFS_AIO */
1783 * syscall - asynchronous write to a file (REALTIME)
1788 sys_aio_write(struct aio_write_args *uap)
1796 error = aio_aqueue(uap->aiocbp, LIO_WRITE);
1799 #endif /* VFS_AIO */
1803 * syscall - XXX undocumented
1808 sys_lio_listio(struct lio_listio_args *uap)
1813 struct proc *p = curproc;
1814 struct lwp *lp = curthread->td_lwp;
1815 int nent, nentqueued;
1816 struct aiocb *iocb, * const *cbptr;
1817 struct aiocblist *cb;
1818 struct kaioinfo *ki;
1819 struct aio_liojob *lj;
1820 int error, runningcode;
1824 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1828 if (nent > AIO_LISTIO_MAX)
1833 if (p->p_aioinfo == NULL)
1834 aio_init_aioinfo(p);
1836 if ((nent + num_queue_count) > max_queue_count) {
1842 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count) {
1847 lj = zalloc(aiolio_zone);
1854 lj->lioj_buffer_count = 0;
1855 lj->lioj_buffer_finished_count = 0;
1856 lj->lioj_queue_count = 0;
1857 lj->lioj_queue_finished_count = 0;
1863 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1864 error = copyin(uap->sig, &lj->lioj_signal,
1865 sizeof(lj->lioj_signal));
1867 zfree(aiolio_zone, lj);
1870 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1871 zfree(aiolio_zone, lj);
1875 lj->lioj_flags |= LIOJ_SIGNAL;
1876 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1878 lj->lioj_flags &= ~LIOJ_SIGNAL;
1880 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1882 * Get pointers to the list of I/O requests.
1886 cbptr = uap->acb_list;
1887 for (i = 0; i < uap->nent; i++) {
1888 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1889 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
1890 error = _aio_aqueue(iocb, lj, 0);
1899 * If we haven't queued any, then just return error.
1901 if (nentqueued == 0) {
1907 * Calculate the appropriate error return.
1913 if (uap->mode == LIO_WAIT) {
1914 int command, found, jobref;
1918 for (i = 0; i < uap->nent; i++) {
1920 * Fetch address of the control buf pointer in
1923 iocb = (struct aiocb *)
1924 (intptr_t)fuword(&cbptr[i]);
1925 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1930 * Fetch the associated command from user space.
1932 command = fuword(&iocb->aio_lio_opcode);
1933 if (command == LIO_NOP) {
1938 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1940 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1941 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1943 if (cb->uaiocb.aio_lio_opcode
1945 lp->lwp_ru.ru_oublock +=
1947 cb->outputcharge = 0;
1948 } else if (cb->uaiocb.aio_lio_opcode
1950 lp->lwp_ru.ru_inblock +=
1952 cb->inputcharge = 0;
1960 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1961 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1971 * If all I/Os have been disposed of, then we can
1974 if (found == nentqueued) {
1975 error = runningcode;
1979 ki->kaio_flags |= KAIO_WAKEUP;
1980 error = tsleep(p, PCATCH, "aiospn", 0);
1982 if (error == EINTR) {
1984 } else if (error == EWOULDBLOCK) {
1991 error = runningcode;
1995 #endif /* VFS_AIO */
2000 * This is a weird hack so that we can post a signal. It is safe to do so from
2001 * a timeout routine, but *not* from an interrupt routine.
2004 process_signal(void *aioj)
2006 struct aiocblist *aiocbe = aioj;
2007 struct aio_liojob *lj = aiocbe->lio;
2008 struct aiocb *cb = &aiocbe->uaiocb;
2010 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2011 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2012 ksignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2013 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2016 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2017 ksignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2021 * Interrupt handler for physio, performs the necessary process wakeups, and
2025 aio_physwakeup(struct bio *bio)
2027 struct buf *bp = bio->bio_buf;
2028 struct aiocblist *aiocbe;
2030 struct kaioinfo *ki;
2031 struct aio_liojob *lj;
2033 aiocbe = bio->bio_caller_info2.ptr;
2037 p = bio->bio_caller_info1.ptr;
2039 aiocbe->jobstate = JOBST_JOBBFINISHED;
2040 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2041 aiocbe->uaiocb._aiocb_private.error = 0;
2042 aiocbe->jobflags |= AIOCBLIST_DONE;
2044 if (bp->b_flags & B_ERROR)
2045 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2049 lj->lioj_buffer_finished_count++;
2052 * wakeup/signal if all of the interrupt jobs are done.
2054 if (lj->lioj_buffer_finished_count ==
2055 lj->lioj_buffer_count) {
2057 * Post a signal if it is called for.
2059 if ((lj->lioj_flags &
2060 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2062 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2063 callout_reset(&aiocbe->timeout, 0,
2064 process_signal, aiocbe);
2071 ki->kaio_buffer_finished_count++;
2072 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2073 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2074 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2076 KNOTE(&aiocbe->klist, 0);
2077 /* Do the wakeup. */
2078 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2079 ki->kaio_flags &= ~KAIO_WAKEUP;
2084 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2085 callout_reset(&aiocbe->timeout, 0,
2086 process_signal, aiocbe);
2092 #endif /* VFS_AIO */
2095 * syscall - wait for the next completion of an aio request
2100 sys_aio_waitcomplete(struct aio_waitcomplete_args *uap)
2105 struct proc *p = curproc;
2106 struct lwp *lp = curthread->td_lwp;
2109 struct kaioinfo *ki;
2110 struct aiocblist *cb = NULL;
2113 suword(uap->aiocbp, (int)NULL);
2117 /* Get timespec struct. */
2118 error = copyin(uap->timeout, &ts, sizeof(ts));
2122 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2125 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2126 if (itimerfix(&atv))
2128 timo = tvtohz_high(&atv);
2138 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2139 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2140 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2141 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2142 lp->lwp_ru.ru_oublock +=
2144 cb->outputcharge = 0;
2145 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2146 lp->lwp_ru.ru_inblock += cb->inputcharge;
2147 cb->inputcharge = 0;
2150 error = cb->uaiocb._aiocb_private.error;
2155 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2157 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
2158 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
2160 error = cb->uaiocb._aiocb_private.error;
2164 ki->kaio_flags |= KAIO_WAKEUP;
2165 error = tsleep(p, PCATCH, "aiowc", timo);
2168 if (error == ERESTART) {
2176 if (error == EWOULDBLOCK) {
2183 #endif /* VFS_AIO */
2188 filt_aioattach(struct knote *kn)
2194 struct filterops aio_filtops =
2195 { 0, filt_aioattach, NULL, NULL };
2198 /* kqueue attach function */
2200 filt_aioattach(struct knote *kn)
2202 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2205 * The aiocbe pointer must be validated before using it, so
2206 * registration is restricted to the kernel; the user cannot
2209 if ((kn->kn_flags & EV_FLAG1) == 0)
2211 kn->kn_flags &= ~EV_FLAG1;
2213 knote_insert(&aiocbe->klist, kn);
2218 /* kqueue detach function */
2220 filt_aiodetach(struct knote *kn)
2222 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2224 knote_remove(&aiocbe->klist, kn);
2227 /* kqueue filter function */
2230 filt_aio(struct knote *kn, long hint)
2232 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2234 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2235 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2236 aiocbe->jobstate != JOBST_JOBBFINISHED)
2238 kn->kn_flags |= EV_EOF;
2242 struct filterops aio_filtops =
2243 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2244 #endif /* VFS_AIO */