Merge from vendor branch SENDMAIL:
[dragonfly.git] / sys / kern / vfs_aio.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1997 John S. Dyson. All rights reserved.
3 *
4 * Redistribution and use in source and binary forms, with or without
5 * modification, are permitted provided that the following conditions
6 * are met:
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.
11 *
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.
15 *
16 * $FreeBSD: src/sys/kern/vfs_aio.c,v 1.70.2.28 2003/05/29 06:15:35 alc Exp $
a5eb27b6 17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.14 2004/09/16 04:42:56 dillon Exp $
984263bc
MD
18 */
19
20/*
21 * This file contains support for the POSIX 1003.1B AIO/LIO facility.
22 */
23
24#include <sys/param.h>
25#include <sys/systm.h>
26#include <sys/buf.h>
27#include <sys/sysproto.h>
28#include <sys/filedesc.h>
29#include <sys/kernel.h>
30#include <sys/fcntl.h>
31#include <sys/file.h>
32#include <sys/lock.h>
33#include <sys/unistd.h>
34#include <sys/proc.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>
41#include <sys/conf.h>
42#include <sys/event.h>
43
44#include <vm/vm.h>
45#include <vm/vm_extern.h>
46#include <vm/pmap.h>
47#include <vm/vm_map.h>
48#include <vm/vm_zone.h>
49#include <sys/aio.h>
7b95be2a
MD
50#include <sys/file2.h>
51#include <sys/buf2.h>
984263bc
MD
52
53#include <machine/limits.h>
54#include "opt_vfs_aio.h"
55
56#ifdef VFS_AIO
57
58/*
59 * Counter for allocating reference ids to new jobs. Wrapped to 1 on
60 * overflow.
61 */
62static long jobrefid;
63
64#define JOBST_NULL 0x0
65#define JOBST_JOBQGLOBAL 0x2
66#define JOBST_JOBRUNNING 0x3
67#define JOBST_JOBFINISHED 0x4
68#define JOBST_JOBQBUF 0x5
69#define JOBST_JOBBFINISHED 0x6
70
71#ifndef MAX_AIO_PER_PROC
72#define MAX_AIO_PER_PROC 32
73#endif
74
75#ifndef MAX_AIO_QUEUE_PER_PROC
76#define MAX_AIO_QUEUE_PER_PROC 256 /* Bigger than AIO_LISTIO_MAX */
77#endif
78
79#ifndef MAX_AIO_PROCS
80#define MAX_AIO_PROCS 32
81#endif
82
83#ifndef MAX_AIO_QUEUE
84#define MAX_AIO_QUEUE 1024 /* Bigger than AIO_LISTIO_MAX */
85#endif
86
87#ifndef TARGET_AIO_PROCS
88#define TARGET_AIO_PROCS 4
89#endif
90
91#ifndef MAX_BUF_AIO
92#define MAX_BUF_AIO 16
93#endif
94
95#ifndef AIOD_TIMEOUT_DEFAULT
96#define AIOD_TIMEOUT_DEFAULT (10 * hz)
97#endif
98
99#ifndef AIOD_LIFETIME_DEFAULT
100#define AIOD_LIFETIME_DEFAULT (30 * hz)
101#endif
102
103SYSCTL_NODE(_vfs, OID_AUTO, aio, CTLFLAG_RW, 0, "Async IO management");
104
105static int max_aio_procs = MAX_AIO_PROCS;
106SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_procs,
107 CTLFLAG_RW, &max_aio_procs, 0,
108 "Maximum number of kernel threads to use for handling async IO");
109
110static int num_aio_procs = 0;
111SYSCTL_INT(_vfs_aio, OID_AUTO, num_aio_procs,
112 CTLFLAG_RD, &num_aio_procs, 0,
113 "Number of presently active kernel threads for async IO");
114
115/*
116 * The code will adjust the actual number of AIO processes towards this
117 * number when it gets a chance.
118 */
119static int target_aio_procs = TARGET_AIO_PROCS;
120SYSCTL_INT(_vfs_aio, OID_AUTO, target_aio_procs, CTLFLAG_RW, &target_aio_procs,
121 0, "Preferred number of ready kernel threads for async IO");
122
123static int max_queue_count = MAX_AIO_QUEUE;
124SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue, CTLFLAG_RW, &max_queue_count, 0,
125 "Maximum number of aio requests to queue, globally");
126
127static int num_queue_count = 0;
128SYSCTL_INT(_vfs_aio, OID_AUTO, num_queue_count, CTLFLAG_RD, &num_queue_count, 0,
129 "Number of queued aio requests");
130
131static int num_buf_aio = 0;
132SYSCTL_INT(_vfs_aio, OID_AUTO, num_buf_aio, CTLFLAG_RD, &num_buf_aio, 0,
133 "Number of aio requests presently handled by the buf subsystem");
134
135/* Number of async I/O thread in the process of being started */
136/* XXX This should be local to _aio_aqueue() */
137static int num_aio_resv_start = 0;
138
139static int aiod_timeout;
140SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_timeout, CTLFLAG_RW, &aiod_timeout, 0,
141 "Timeout value for synchronous aio operations");
142
143static int aiod_lifetime;
144SYSCTL_INT(_vfs_aio, OID_AUTO, aiod_lifetime, CTLFLAG_RW, &aiod_lifetime, 0,
145 "Maximum lifetime for idle aiod");
146
147static int max_aio_per_proc = MAX_AIO_PER_PROC;
148SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_per_proc, CTLFLAG_RW, &max_aio_per_proc,
149 0, "Maximum active aio requests per process (stored in the process)");
150
151static int max_aio_queue_per_proc = MAX_AIO_QUEUE_PER_PROC;
152SYSCTL_INT(_vfs_aio, OID_AUTO, max_aio_queue_per_proc, CTLFLAG_RW,
153 &max_aio_queue_per_proc, 0,
154 "Maximum queued aio requests per process (stored in the process)");
155
156static int max_buf_aio = MAX_BUF_AIO;
157SYSCTL_INT(_vfs_aio, OID_AUTO, max_buf_aio, CTLFLAG_RW, &max_buf_aio, 0,
158 "Maximum buf aio requests per process (stored in the process)");
159
160/*
161 * AIO process info
162 */
163#define AIOP_FREE 0x1 /* proc on free queue */
164#define AIOP_SCHED 0x2 /* proc explicitly scheduled */
165
166struct aioproclist {
167 int aioprocflags; /* AIO proc flags */
168 TAILQ_ENTRY(aioproclist) list; /* List of processes */
169 struct proc *aioproc; /* The AIO thread */
170};
171
172/*
173 * data-structure for lio signal management
174 */
175struct aio_liojob {
176 int lioj_flags;
177 int lioj_buffer_count;
178 int lioj_buffer_finished_count;
179 int lioj_queue_count;
180 int lioj_queue_finished_count;
181 struct sigevent lioj_signal; /* signal on all I/O done */
182 TAILQ_ENTRY(aio_liojob) lioj_list;
183 struct kaioinfo *lioj_ki;
184};
185#define LIOJ_SIGNAL 0x1 /* signal on all done (lio) */
186#define LIOJ_SIGNAL_POSTED 0x2 /* signal has been posted */
187
188/*
189 * per process aio data structure
190 */
191struct kaioinfo {
192 int kaio_flags; /* per process kaio flags */
193 int kaio_maxactive_count; /* maximum number of AIOs */
194 int kaio_active_count; /* number of currently used AIOs */
195 int kaio_qallowed_count; /* maxiumu size of AIO queue */
196 int kaio_queue_count; /* size of AIO queue */
197 int kaio_ballowed_count; /* maximum number of buffers */
198 int kaio_queue_finished_count; /* number of daemon jobs finished */
199 int kaio_buffer_count; /* number of physio buffers */
200 int kaio_buffer_finished_count; /* count of I/O done */
201 struct proc *kaio_p; /* process that uses this kaio block */
202 TAILQ_HEAD(,aio_liojob) kaio_liojoblist; /* list of lio jobs */
203 TAILQ_HEAD(,aiocblist) kaio_jobqueue; /* job queue for process */
204 TAILQ_HEAD(,aiocblist) kaio_jobdone; /* done queue for process */
205 TAILQ_HEAD(,aiocblist) kaio_bufqueue; /* buffer job queue for process */
206 TAILQ_HEAD(,aiocblist) kaio_bufdone; /* buffer done queue for process */
207 TAILQ_HEAD(,aiocblist) kaio_sockqueue; /* queue for aios waiting on sockets */
208};
209
210#define KAIO_RUNDOWN 0x1 /* process is being run down */
211#define KAIO_WAKEUP 0x2 /* wakeup process when there is a significant event */
212
213static TAILQ_HEAD(,aioproclist) aio_freeproc, aio_activeproc;
214static TAILQ_HEAD(,aiocblist) aio_jobs; /* Async job list */
215static TAILQ_HEAD(,aiocblist) aio_bufjobs; /* Phys I/O job list */
216static TAILQ_HEAD(,aiocblist) aio_freejobs; /* Pool of free jobs */
217
218static void aio_init_aioinfo(struct proc *p);
219static void aio_onceonly(void *);
220static int aio_free_entry(struct aiocblist *aiocbe);
221static void aio_process(struct aiocblist *aiocbe);
222static int aio_newproc(void);
41c20dac 223static int aio_aqueue(struct aiocb *job, int type);
984263bc
MD
224static void aio_physwakeup(struct buf *bp);
225static int aio_fphysio(struct aiocblist *aiocbe);
226static int aio_qphysio(struct proc *p, struct aiocblist *iocb);
227static void aio_daemon(void *uproc);
228static void process_signal(void *aioj);
229
230SYSINIT(aio, SI_SUB_VFS, SI_ORDER_ANY, aio_onceonly, NULL);
231
232/*
233 * Zones for:
234 * kaio Per process async io info
235 * aiop async io thread data
236 * aiocb async io jobs
237 * aiol list io job pointer - internal to aio_suspend XXX
238 * aiolio list io jobs
239 */
240static vm_zone_t kaio_zone, aiop_zone, aiocb_zone, aiol_zone, aiolio_zone;
241
242/*
243 * Startup initialization
244 */
245static void
246aio_onceonly(void *na)
247{
248 TAILQ_INIT(&aio_freeproc);
249 TAILQ_INIT(&aio_activeproc);
250 TAILQ_INIT(&aio_jobs);
251 TAILQ_INIT(&aio_bufjobs);
252 TAILQ_INIT(&aio_freejobs);
253 kaio_zone = zinit("AIO", sizeof(struct kaioinfo), 0, 0, 1);
254 aiop_zone = zinit("AIOP", sizeof(struct aioproclist), 0, 0, 1);
255 aiocb_zone = zinit("AIOCB", sizeof(struct aiocblist), 0, 0, 1);
256 aiol_zone = zinit("AIOL", AIO_LISTIO_MAX*sizeof(intptr_t), 0, 0, 1);
257 aiolio_zone = zinit("AIOLIO", sizeof(struct aio_liojob), 0, 0, 1);
258 aiod_timeout = AIOD_TIMEOUT_DEFAULT;
259 aiod_lifetime = AIOD_LIFETIME_DEFAULT;
260 jobrefid = 1;
261}
262
263/*
264 * Init the per-process aioinfo structure. The aioinfo limits are set
265 * per-process for user limit (resource) management.
266 */
267static void
268aio_init_aioinfo(struct proc *p)
269{
270 struct kaioinfo *ki;
271 if (p->p_aioinfo == NULL) {
272 ki = zalloc(kaio_zone);
273 p->p_aioinfo = ki;
274 ki->kaio_flags = 0;
275 ki->kaio_maxactive_count = max_aio_per_proc;
276 ki->kaio_active_count = 0;
277 ki->kaio_qallowed_count = max_aio_queue_per_proc;
278 ki->kaio_queue_count = 0;
279 ki->kaio_ballowed_count = max_buf_aio;
280 ki->kaio_buffer_count = 0;
281 ki->kaio_buffer_finished_count = 0;
282 ki->kaio_p = p;
283 TAILQ_INIT(&ki->kaio_jobdone);
284 TAILQ_INIT(&ki->kaio_jobqueue);
285 TAILQ_INIT(&ki->kaio_bufdone);
286 TAILQ_INIT(&ki->kaio_bufqueue);
287 TAILQ_INIT(&ki->kaio_liojoblist);
288 TAILQ_INIT(&ki->kaio_sockqueue);
289 }
290
291 while (num_aio_procs < target_aio_procs)
292 aio_newproc();
293}
294
295/*
296 * Free a job entry. Wait for completion if it is currently active, but don't
297 * delay forever. If we delay, we return a flag that says that we have to
298 * restart the queue scan.
299 */
300static int
301aio_free_entry(struct aiocblist *aiocbe)
302{
303 struct kaioinfo *ki;
304 struct aio_liojob *lj;
305 struct proc *p;
306 int error;
307 int s;
308
309 if (aiocbe->jobstate == JOBST_NULL)
310 panic("aio_free_entry: freeing already free job");
311
312 p = aiocbe->userproc;
313 ki = p->p_aioinfo;
314 lj = aiocbe->lio;
315 if (ki == NULL)
316 panic("aio_free_entry: missing p->p_aioinfo");
317
318 while (aiocbe->jobstate == JOBST_JOBRUNNING) {
319 aiocbe->jobflags |= AIOCBLIST_RUNDOWN;
377d4740 320 tsleep(aiocbe, 0, "jobwai", 0);
984263bc
MD
321 }
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");
327
328 if (lj) {
329 lj->lioj_queue_count--;
330 if (aiocbe->jobflags & AIOCBLIST_DONE)
331 lj->lioj_queue_finished_count--;
332 }
333 ki->kaio_queue_count--;
334 if (aiocbe->jobflags & AIOCBLIST_DONE)
335 ki->kaio_queue_finished_count--;
336 num_queue_count--;
337 } else {
338 if (lj) {
339 lj->lioj_buffer_count--;
340 if (aiocbe->jobflags & AIOCBLIST_DONE)
341 lj->lioj_buffer_finished_count--;
342 }
343 if (aiocbe->jobflags & AIOCBLIST_DONE)
344 ki->kaio_buffer_finished_count--;
345 ki->kaio_buffer_count--;
346 num_buf_aio--;
347 }
348
349 /* aiocbe is going away, we need to destroy any knotes */
7b95be2a 350 knote_remove(p->p_thread, &aiocbe->klist);
984263bc
MD
351
352 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags & KAIO_RUNDOWN)
353 && ((ki->kaio_buffer_count == 0) && (ki->kaio_queue_count == 0)))) {
354 ki->kaio_flags &= ~KAIO_WAKEUP;
355 wakeup(p);
356 }
357
358 if (aiocbe->jobstate == JOBST_JOBQBUF) {
359 if ((error = aio_fphysio(aiocbe)) != 0)
360 return error;
361 if (aiocbe->jobstate != JOBST_JOBBFINISHED)
362 panic("aio_free_entry: invalid physio finish-up state");
363 s = splbio();
364 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
365 splx(s);
366 } else if (aiocbe->jobstate == JOBST_JOBQGLOBAL) {
367 s = splnet();
368 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
369 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
370 splx(s);
371 } else if (aiocbe->jobstate == JOBST_JOBFINISHED)
372 TAILQ_REMOVE(&ki->kaio_jobdone, aiocbe, plist);
373 else if (aiocbe->jobstate == JOBST_JOBBFINISHED) {
374 s = splbio();
375 TAILQ_REMOVE(&ki->kaio_bufdone, aiocbe, plist);
376 splx(s);
377 if (aiocbe->bp) {
378 vunmapbuf(aiocbe->bp);
379 relpbuf(aiocbe->bp, NULL);
380 aiocbe->bp = NULL;
381 }
382 }
383 if (lj && (lj->lioj_buffer_count == 0) && (lj->lioj_queue_count == 0)) {
384 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
385 zfree(aiolio_zone, lj);
386 }
387 aiocbe->jobstate = JOBST_NULL;
a5eb27b6 388 callout_stop(&aiocbe->timeout);
7b95be2a 389 fdrop(aiocbe->fd_file, curthread);
984263bc
MD
390 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
391 return 0;
392}
393#endif /* VFS_AIO */
394
395/*
396 * Rundown the jobs for a given process.
397 */
398void
399aio_proc_rundown(struct proc *p)
400{
401#ifndef VFS_AIO
402 return;
403#else
404 int s;
405 struct kaioinfo *ki;
406 struct aio_liojob *lj, *ljn;
407 struct aiocblist *aiocbe, *aiocbn;
408 struct file *fp;
409 struct socket *so;
410
411 ki = p->p_aioinfo;
412 if (ki == NULL)
413 return;
414
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;
377d4740 419 if (tsleep(p, 0, "kaiowt", aiod_timeout))
984263bc
MD
420 break;
421 }
422
423 /*
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.
426 */
427 s = splnet();
428 for (aiocbe = TAILQ_FIRST(&ki->kaio_sockqueue); aiocbe; aiocbe =
429 aiocbn) {
430 aiocbn = TAILQ_NEXT(aiocbe, plist);
431 fp = aiocbe->fd_file;
432 if (fp != NULL) {
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;
438 }
439 }
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);
443 }
444 splx(s);
445
446restart1:
447 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobdone); aiocbe; aiocbe = aiocbn) {
448 aiocbn = TAILQ_NEXT(aiocbe, plist);
449 if (aio_free_entry(aiocbe))
450 goto restart1;
451 }
452
453restart2:
454 for (aiocbe = TAILQ_FIRST(&ki->kaio_jobqueue); aiocbe; aiocbe =
455 aiocbn) {
456 aiocbn = TAILQ_NEXT(aiocbe, plist);
457 if (aio_free_entry(aiocbe))
458 goto restart2;
459 }
460
461/*
462 * Note the use of lots of splbio here, trying to avoid splbio for long chains
463 * of I/O. Probably unnecessary.
464 */
465restart3:
466 s = splbio();
467 while (TAILQ_FIRST(&ki->kaio_bufqueue)) {
468 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 469 tsleep(p, 0, "aioprn", 0);
984263bc
MD
470 splx(s);
471 goto restart3;
472 }
473 splx(s);
474
475restart4:
476 s = splbio();
477 for (aiocbe = TAILQ_FIRST(&ki->kaio_bufdone); aiocbe; aiocbe = aiocbn) {
478 aiocbn = TAILQ_NEXT(aiocbe, plist);
479 if (aio_free_entry(aiocbe)) {
480 splx(s);
481 goto restart4;
482 }
483 }
484 splx(s);
485
486 /*
487 * If we've slept, jobs might have moved from one queue to another.
488 * Retry rundown if we didn't manage to empty the queues.
489 */
490 if (TAILQ_FIRST(&ki->kaio_jobdone) != NULL ||
491 TAILQ_FIRST(&ki->kaio_jobqueue) != NULL ||
492 TAILQ_FIRST(&ki->kaio_bufqueue) != NULL ||
493 TAILQ_FIRST(&ki->kaio_bufdone) != NULL)
494 goto restart1;
495
496 for (lj = TAILQ_FIRST(&ki->kaio_liojoblist); lj; lj = ljn) {
497 ljn = TAILQ_NEXT(lj, lioj_list);
498 if ((lj->lioj_buffer_count == 0) && (lj->lioj_queue_count ==
499 0)) {
500 TAILQ_REMOVE(&ki->kaio_liojoblist, lj, lioj_list);
501 zfree(aiolio_zone, lj);
502 } else {
503#ifdef DIAGNOSTIC
504 printf("LIO job not cleaned up: B:%d, BF:%d, Q:%d, "
505 "QF:%d\n", lj->lioj_buffer_count,
506 lj->lioj_buffer_finished_count,
507 lj->lioj_queue_count,
508 lj->lioj_queue_finished_count);
509#endif
510 }
511 }
512
513 zfree(kaio_zone, ki);
514 p->p_aioinfo = NULL;
515#endif /* VFS_AIO */
516}
517
518#ifdef VFS_AIO
519/*
520 * Select a job to run (called by an AIO daemon).
521 */
522static struct aiocblist *
523aio_selectjob(struct aioproclist *aiop)
524{
525 int s;
526 struct aiocblist *aiocbe;
527 struct kaioinfo *ki;
528 struct proc *userp;
529
530 s = splnet();
531 for (aiocbe = TAILQ_FIRST(&aio_jobs); aiocbe; aiocbe =
532 TAILQ_NEXT(aiocbe, list)) {
533 userp = aiocbe->userproc;
534 ki = userp->p_aioinfo;
535
536 if (ki->kaio_active_count < ki->kaio_maxactive_count) {
537 TAILQ_REMOVE(&aio_jobs, aiocbe, list);
538 splx(s);
539 return aiocbe;
540 }
541 }
542 splx(s);
543
544 return NULL;
545}
546
547/*
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.
552 */
553static void
554aio_process(struct aiocblist *aiocbe)
555{
7b95be2a 556 struct thread *mytd;
984263bc
MD
557 struct aiocb *cb;
558 struct file *fp;
559 struct uio auio;
560 struct iovec aiov;
561 int cnt;
562 int error;
563 int oublock_st, oublock_end;
564 int inblock_st, inblock_end;
565
7b95be2a 566 mytd = curthread;
984263bc
MD
567 cb = &aiocbe->uaiocb;
568 fp = aiocbe->fd_file;
569
570 aiov.iov_base = (void *)(uintptr_t)cb->aio_buf;
571 aiov.iov_len = cb->aio_nbytes;
572
573 auio.uio_iov = &aiov;
574 auio.uio_iovcnt = 1;
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;
7b95be2a 579 auio.uio_td = mytd;
984263bc 580
7b95be2a
MD
581 inblock_st = mytd->td_proc->p_stats->p_ru.ru_inblock;
582 oublock_st = mytd->td_proc->p_stats->p_ru.ru_oublock;
984263bc
MD
583 /*
584 * _aio_aqueue() acquires a reference to the file that is
585 * released in aio_free_entry().
586 */
587 if (cb->aio_lio_opcode == LIO_READ) {
588 auio.uio_rw = UIO_READ;
7b95be2a 589 error = fo_read(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
984263bc
MD
590 } else {
591 auio.uio_rw = UIO_WRITE;
7b95be2a 592 error = fo_write(fp, &auio, fp->f_cred, FOF_OFFSET, mytd);
984263bc 593 }
7b95be2a
MD
594 inblock_end = mytd->td_proc->p_stats->p_ru.ru_inblock;
595 oublock_end = mytd->td_proc->p_stats->p_ru.ru_oublock;
984263bc
MD
596
597 aiocbe->inputcharge = inblock_end - inblock_st;
598 aiocbe->outputcharge = oublock_end - oublock_st;
599
600 if ((error) && (auio.uio_resid != cnt)) {
601 if (error == ERESTART || error == EINTR || error == EWOULDBLOCK)
602 error = 0;
603 if ((error == EPIPE) && (cb->aio_lio_opcode == LIO_WRITE))
604 psignal(aiocbe->userproc, SIGPIPE);
605 }
606
607 cnt -= auio.uio_resid;
608 cb->_aiocb_private.error = error;
609 cb->_aiocb_private.status = cnt;
610}
611
612/*
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.
8a8d5d85
MD
615 *
616 * The MP lock is held on entry.
984263bc
MD
617 */
618static void
619aio_daemon(void *uproc)
620{
621 int s;
622 struct aio_liojob *lj;
623 struct aiocb *cb;
624 struct aiocblist *aiocbe;
625 struct aioproclist *aiop;
626 struct kaioinfo *ki;
627 struct proc *curcp, *mycp, *userp;
628 struct vmspace *myvm, *tmpvm;
e9a372eb 629 struct ucred *cr;
984263bc
MD
630
631 /*
632 * Local copies of curproc (cp) and vmspace (myvm)
633 */
634 mycp = curproc;
635 myvm = mycp->p_vmspace;
636
637 if (mycp->p_textvp) {
638 vrele(mycp->p_textvp);
639 mycp->p_textvp = NULL;
640 }
641
642 /*
643 * Allocate and ready the aio control info. There is one aiop structure
644 * per daemon.
645 */
646 aiop = zalloc(aiop_zone);
647 aiop->aioproc = mycp;
648 aiop->aioprocflags |= AIOP_FREE;
649
650 s = splnet();
651
652 /*
653 * Place thread (lightweight process) onto the AIO free thread list.
654 */
655 if (TAILQ_EMPTY(&aio_freeproc))
656 wakeup(&aio_freeproc);
657 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
658
659 splx(s);
660
661 /* Make up a name for the daemon. */
662 strcpy(mycp->p_comm, "aiod");
663
664 /*
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".
668 */
669 fdfree(mycp);
670 mycp->p_fd = NULL;
e9a372eb
MD
671 cr = cratom(&mycp->p_ucred);
672 cr->cr_uid = 0;
6eedb489 673 uireplace(&cr->cr_uidinfo, uifind(0));
e9a372eb
MD
674 cr->cr_ngroups = 1;
675 cr->cr_groups[0] = 1;
984263bc
MD
676
677 /* The daemon resides in its own pgrp. */
678 enterpgrp(mycp, mycp->p_pid, 1);
679
680 /* Mark special process type. */
681 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
682
683 /*
684 * Wakeup parent process. (Parent sleeps to keep from blasting away
685 * and creating too many daemons.)
686 */
687 wakeup(mycp);
688
689 for (;;) {
690 /*
691 * curcp is the current daemon process context.
692 * userp is the current user process context.
693 */
694 curcp = mycp;
695
696 /*
697 * Take daemon off of free queue
698 */
699 if (aiop->aioprocflags & AIOP_FREE) {
700 s = splnet();
701 TAILQ_REMOVE(&aio_freeproc, aiop, list);
702 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
703 aiop->aioprocflags &= ~AIOP_FREE;
704 splx(s);
705 }
706 aiop->aioprocflags &= ~AIOP_SCHED;
707
708 /*
709 * Check for jobs.
710 */
711 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
712 cb = &aiocbe->uaiocb;
713 userp = aiocbe->userproc;
714
715 aiocbe->jobstate = JOBST_JOBRUNNING;
716
717 /*
718 * Connect to process address space for user program.
719 */
720 if (userp != curcp) {
721 /*
722 * Save the current address space that we are
723 * connected to.
724 */
725 tmpvm = mycp->p_vmspace;
726
727 /*
728 * Point to the new user address space, and
729 * refer to it.
730 */
731 mycp->p_vmspace = userp->p_vmspace;
732 mycp->p_vmspace->vm_refcnt++;
733
734 /* Activate the new mapping. */
735 pmap_activate(mycp);
736
737 /*
738 * If the old address space wasn't the daemons
739 * own address space, then we need to remove the
740 * daemon's reference from the other process
741 * that it was acting on behalf of.
742 */
743 if (tmpvm != myvm) {
744 vmspace_free(tmpvm);
745 }
746 curcp = userp;
747 }
748
749 ki = userp->p_aioinfo;
750 lj = aiocbe->lio;
751
752 /* Account for currently active jobs. */
753 ki->kaio_active_count++;
754
755 /* Do the I/O function. */
756 aio_process(aiocbe);
757
758 /* Decrement the active job count. */
759 ki->kaio_active_count--;
760
761 /*
762 * Increment the completion count for wakeup/signal
763 * comparisons.
764 */
765 aiocbe->jobflags |= AIOCBLIST_DONE;
766 ki->kaio_queue_finished_count++;
767 if (lj)
768 lj->lioj_queue_finished_count++;
769 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
770 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
771 ki->kaio_flags &= ~KAIO_WAKEUP;
772 wakeup(userp);
773 }
774
775 s = splbio();
776 if (lj && (lj->lioj_flags &
777 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
778 if ((lj->lioj_queue_finished_count ==
779 lj->lioj_queue_count) &&
780 (lj->lioj_buffer_finished_count ==
781 lj->lioj_buffer_count)) {
782 psignal(userp,
783 lj->lioj_signal.sigev_signo);
784 lj->lioj_flags |=
785 LIOJ_SIGNAL_POSTED;
786 }
787 }
788 splx(s);
789
790 aiocbe->jobstate = JOBST_JOBFINISHED;
791
792 s = splnet();
793 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
794 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
795 splx(s);
796 KNOTE(&aiocbe->klist, 0);
797
798 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
799 wakeup(aiocbe);
800 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
801 }
802
803 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
804 psignal(userp, cb->aio_sigevent.sigev_signo);
805 }
806 }
807
808 /*
809 * Disconnect from user address space.
810 */
811 if (curcp != mycp) {
812 /* Get the user address space to disconnect from. */
813 tmpvm = mycp->p_vmspace;
814
815 /* Get original address space for daemon. */
816 mycp->p_vmspace = myvm;
817
818 /* Activate the daemon's address space. */
819 pmap_activate(mycp);
820#ifdef DIAGNOSTIC
821 if (tmpvm == myvm) {
822 printf("AIOD: vmspace problem -- %d\n",
823 mycp->p_pid);
824 }
825#endif
826 /* Remove our vmspace reference. */
827 vmspace_free(tmpvm);
828
829 curcp = mycp;
830 }
831
832 /*
833 * If we are the first to be put onto the free queue, wakeup
834 * anyone waiting for a daemon.
835 */
836 s = splnet();
837 TAILQ_REMOVE(&aio_activeproc, aiop, list);
838 if (TAILQ_EMPTY(&aio_freeproc))
839 wakeup(&aio_freeproc);
840 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
841 aiop->aioprocflags |= AIOP_FREE;
842 splx(s);
843
844 /*
845 * If daemon is inactive for a long time, allow it to exit,
846 * thereby freeing resources.
847 */
848 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
377d4740 849 0, "aiordy", aiod_lifetime)) {
984263bc
MD
850 s = splnet();
851 if (TAILQ_EMPTY(&aio_jobs)) {
852 if ((aiop->aioprocflags & AIOP_FREE) &&
853 (num_aio_procs > target_aio_procs)) {
854 TAILQ_REMOVE(&aio_freeproc, aiop, list);
855 splx(s);
856 zfree(aiop_zone, aiop);
857 num_aio_procs--;
858#ifdef DIAGNOSTIC
859 if (mycp->p_vmspace->vm_refcnt <= 1) {
860 printf("AIOD: bad vm refcnt for"
861 " exiting daemon: %d\n",
862 mycp->p_vmspace->vm_refcnt);
863 }
864#endif
7b95be2a 865 exit1(0);
984263bc
MD
866 }
867 }
868 splx(s);
869 }
870 }
871}
872
873/*
874 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
875 * AIO daemon modifies its environment itself.
876 */
877static int
878aio_newproc()
879{
880 int error;
881 struct proc *p, *np;
882
883 p = &proc0;
884 error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
885 if (error)
886 return error;
887 cpu_set_fork_handler(np, aio_daemon, curproc);
7d0bac62 888 start_forked_proc(p, np);
984263bc
MD
889
890 /*
891 * Wait until daemon is started, but continue on just in case to
892 * handle error conditions.
893 */
377d4740 894 error = tsleep(np, 0, "aiosta", aiod_timeout);
984263bc
MD
895 num_aio_procs++;
896
897 return error;
898}
899
900/*
901 * Try the high-performance, low-overhead physio method for eligible
902 * VCHR devices. This method doesn't use an aio helper thread, and
903 * thus has very low overhead.
904 *
905 * Assumes that the caller, _aio_aqueue(), has incremented the file
906 * structure's reference count, preventing its deallocation for the
907 * duration of this call.
908 */
909static int
910aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
911{
912 int error;
913 struct aiocb *cb;
914 struct file *fp;
915 struct buf *bp;
916 struct vnode *vp;
917 struct kaioinfo *ki;
918 struct aio_liojob *lj;
919 int s;
920 int notify;
921
922 cb = &aiocbe->uaiocb;
923 fp = aiocbe->fd_file;
924
925 if (fp->f_type != DTYPE_VNODE)
926 return (-1);
927
928 vp = (struct vnode *)fp->f_data;
929
930 /*
931 * If its not a disk, we don't want to return a positive error.
932 * It causes the aio code to not fall through to try the thread
933 * way when you're talking to a regular file.
934 */
935 if (!vn_isdisk(vp, &error)) {
936 if (error == ENOTBLK)
937 return (-1);
938 else
939 return (error);
940 }
941
942 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
943 return (-1);
944
945 if (cb->aio_nbytes >
946 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
947 return (-1);
948
949 ki = p->p_aioinfo;
950 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
951 return (-1);
952
953 ki->kaio_buffer_count++;
954
955 lj = aiocbe->lio;
956 if (lj)
957 lj->lioj_buffer_count++;
958
959 /* Create and build a buffer header for a transfer. */
960 bp = (struct buf *)getpbuf(NULL);
961 BUF_KERNPROC(bp);
962
963 /*
964 * Get a copy of the kva from the physical buffer.
965 */
966 bp->b_caller1 = p;
967 bp->b_dev = vp->v_rdev;
968 error = 0;
969
970 bp->b_bcount = cb->aio_nbytes;
971 bp->b_bufsize = cb->aio_nbytes;
972 bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
973 B_WRITE : B_READ);
974 bp->b_iodone = aio_physwakeup;
975 bp->b_saveaddr = bp->b_data;
976 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
977 bp->b_blkno = btodb(cb->aio_offset);
978
979 /* Bring buffer into kernel space. */
980 if (vmapbuf(bp) < 0) {
981 error = EFAULT;
982 goto doerror;
983 }
984
985 s = splbio();
986 aiocbe->bp = bp;
987 bp->b_spc = (void *)aiocbe;
988 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
989 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
990 aiocbe->jobstate = JOBST_JOBQBUF;
991 cb->_aiocb_private.status = cb->aio_nbytes;
992 num_buf_aio++;
993 bp->b_error = 0;
994
995 splx(s);
996
997 /* Perform transfer. */
998 BUF_STRATEGY(bp, 0);
999
1000 notify = 0;
1001 s = splbio();
1002
1003 /*
1004 * If we had an error invoking the request, or an error in processing
1005 * the request before we have returned, we process it as an error in
1006 * transfer. Note that such an I/O error is not indicated immediately,
1007 * but is returned using the aio_error mechanism. In this case,
1008 * aio_suspend will return immediately.
1009 */
1010 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1011 struct aiocb *job = aiocbe->uuaiocb;
1012
1013 aiocbe->uaiocb._aiocb_private.status = 0;
1014 suword(&job->_aiocb_private.status, 0);
1015 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1016 suword(&job->_aiocb_private.error, bp->b_error);
1017
1018 ki->kaio_buffer_finished_count++;
1019
1020 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1021 aiocbe->jobstate = JOBST_JOBBFINISHED;
1022 aiocbe->jobflags |= AIOCBLIST_DONE;
1023 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1024 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1025 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1026 notify = 1;
1027 }
1028 }
1029 splx(s);
1030 if (notify)
1031 KNOTE(&aiocbe->klist, 0);
1032 return 0;
1033
1034doerror:
1035 ki->kaio_buffer_count--;
1036 if (lj)
1037 lj->lioj_buffer_count--;
1038 aiocbe->bp = NULL;
1039 relpbuf(bp, NULL);
1040 return error;
1041}
1042
1043/*
1044 * This waits/tests physio completion.
1045 */
1046static int
1047aio_fphysio(struct aiocblist *iocb)
1048{
1049 int s;
1050 struct buf *bp;
1051 int error;
1052
1053 bp = iocb->bp;
1054
1055 s = splbio();
1056 while ((bp->b_flags & B_DONE) == 0) {
377d4740 1057 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
984263bc
MD
1058 if ((bp->b_flags & B_DONE) == 0) {
1059 splx(s);
1060 return EINPROGRESS;
1061 } else
1062 break;
1063 }
1064 }
1065 splx(s);
1066
1067 /* Release mapping into kernel space. */
1068 vunmapbuf(bp);
1069 iocb->bp = 0;
1070
1071 error = 0;
1072
1073 /* Check for an error. */
1074 if (bp->b_flags & B_ERROR)
1075 error = bp->b_error;
1076
1077 relpbuf(bp, NULL);
1078 return (error);
1079}
1080#endif /* VFS_AIO */
1081
1082/*
1083 * Wake up aio requests that may be serviceable now.
1084 */
1085void
1086aio_swake(struct socket *so, struct sockbuf *sb)
1087{
1088#ifndef VFS_AIO
1089 return;
1090#else
1091 struct aiocblist *cb,*cbn;
1092 struct proc *p;
1093 struct kaioinfo *ki = NULL;
1094 int opcode, wakecount = 0;
1095 struct aioproclist *aiop;
1096
1097 if (sb == &so->so_snd) {
1098 opcode = LIO_WRITE;
1099 so->so_snd.sb_flags &= ~SB_AIO;
1100 } else {
1101 opcode = LIO_READ;
1102 so->so_rcv.sb_flags &= ~SB_AIO;
1103 }
1104
1105 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1106 cbn = TAILQ_NEXT(cb, list);
1107 if (opcode == cb->uaiocb.aio_lio_opcode) {
1108 p = cb->userproc;
1109 ki = p->p_aioinfo;
1110 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1111 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1112 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1113 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1114 wakecount++;
1115 if (cb->jobstate != JOBST_JOBQGLOBAL)
1116 panic("invalid queue value");
1117 }
1118 }
1119
1120 while (wakecount--) {
1121 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1122 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1123 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1124 aiop->aioprocflags &= ~AIOP_FREE;
1125 wakeup(aiop->aioproc);
1126 }
1127 }
1128#endif /* VFS_AIO */
1129}
1130
1131#ifdef VFS_AIO
1132/*
1133 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1134 * technique is done in this code.
1135 */
1136static int
41c20dac 1137_aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
984263bc 1138{
7b95be2a 1139 struct proc *p = curproc;
984263bc
MD
1140 struct filedesc *fdp;
1141 struct file *fp;
1142 unsigned int fd;
1143 struct socket *so;
1144 int s;
1145 int error;
1146 int opcode, user_opcode;
1147 struct aiocblist *aiocbe;
1148 struct aioproclist *aiop;
1149 struct kaioinfo *ki;
1150 struct kevent kev;
1151 struct kqueue *kq;
1152 struct file *kq_fp;
1153
1154 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1155 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1156 else
1157 aiocbe = zalloc (aiocb_zone);
1158
1159 aiocbe->inputcharge = 0;
1160 aiocbe->outputcharge = 0;
a5eb27b6 1161 callout_init(&aiocbe->timeout);
984263bc
MD
1162 SLIST_INIT(&aiocbe->klist);
1163
1164 suword(&job->_aiocb_private.status, -1);
1165 suword(&job->_aiocb_private.error, 0);
1166 suword(&job->_aiocb_private.kernelinfo, -1);
1167
1168 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1169 if (error) {
1170 suword(&job->_aiocb_private.error, error);
1171 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1172 return error;
1173 }
1174 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1175 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1176 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1177 return EINVAL;
1178 }
1179
1180 /* Save userspace address of the job info. */
1181 aiocbe->uuaiocb = job;
1182
1183 /* Get the opcode. */
1184 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1185 if (type != LIO_NOP)
1186 aiocbe->uaiocb.aio_lio_opcode = type;
1187 opcode = aiocbe->uaiocb.aio_lio_opcode;
1188
1189 /* Get the fd info for process. */
1190 fdp = p->p_fd;
1191
1192 /*
1193 * Range check file descriptor.
1194 */
1195 fd = aiocbe->uaiocb.aio_fildes;
1196 if (fd >= fdp->fd_nfiles) {
1197 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1198 if (type == 0)
1199 suword(&job->_aiocb_private.error, EBADF);
1200 return EBADF;
1201 }
1202
1203 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1204 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1205 0))) {
1206 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1207 if (type == 0)
1208 suword(&job->_aiocb_private.error, EBADF);
1209 return EBADF;
1210 }
1211 fhold(fp);
1212
1213 if (aiocbe->uaiocb.aio_offset == -1LL) {
1214 error = EINVAL;
1215 goto aqueue_fail;
1216 }
1217 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1218 if (error) {
1219 error = EINVAL;
1220 goto aqueue_fail;
1221 }
1222 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1223 if (jobrefid == LONG_MAX)
1224 jobrefid = 1;
1225 else
1226 jobrefid++;
1227
1228 if (opcode == LIO_NOP) {
7b95be2a 1229 fdrop(fp, p->p_thread);
984263bc
MD
1230 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1231 if (type == 0) {
1232 suword(&job->_aiocb_private.error, 0);
1233 suword(&job->_aiocb_private.status, 0);
1234 suword(&job->_aiocb_private.kernelinfo, 0);
1235 }
1236 return 0;
1237 }
1238 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1239 if (type == 0)
1240 suword(&job->_aiocb_private.status, 0);
1241 error = EINVAL;
1242 goto aqueue_fail;
1243 }
1244
1245 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1246 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1247 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1248 }
1249 else {
1250 /*
1251 * This method for requesting kevent-based notification won't
1252 * work on the alpha, since we're passing in a pointer
1253 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1254 * based method instead.
1255 */
1256 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1257 user_opcode == LIO_WRITE)
1258 goto no_kqueue;
1259
1260 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1261 &kev, sizeof(kev));
1262 if (error)
1263 goto aqueue_fail;
1264 }
1265 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1266 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1267 (kq_fp->f_type != DTYPE_KQUEUE)) {
1268 error = EBADF;
1269 goto aqueue_fail;
1270 }
1271 kq = (struct kqueue *)kq_fp->f_data;
1272 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1273 kev.filter = EVFILT_AIO;
1274 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1275 kev.data = (intptr_t)aiocbe;
7b95be2a 1276 error = kqueue_register(kq, &kev, p->p_thread);
984263bc
MD
1277aqueue_fail:
1278 if (error) {
7b95be2a 1279 fdrop(fp, p->p_thread);
984263bc
MD
1280 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1281 if (type == 0)
1282 suword(&job->_aiocb_private.error, error);
1283 goto done;
1284 }
1285no_kqueue:
1286
1287 suword(&job->_aiocb_private.error, EINPROGRESS);
1288 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1289 aiocbe->userproc = p;
1290 aiocbe->jobflags = 0;
1291 aiocbe->lio = lj;
1292 ki = p->p_aioinfo;
1293
1294 if (fp->f_type == DTYPE_SOCKET) {
1295 /*
1296 * Alternate queueing for socket ops: Reach down into the
1297 * descriptor to get the socket data. Then check to see if the
1298 * socket is ready to be read or written (based on the requested
1299 * operation).
1300 *
1301 * If it is not ready for io, then queue the aiocbe on the
1302 * socket, and set the flags so we get a call when sbnotify()
1303 * happens.
1304 */
1305 so = (struct socket *)fp->f_data;
1306 s = splnet();
1307 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1308 LIO_WRITE) && (!sowriteable(so)))) {
1309 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1310 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1311 if (opcode == LIO_READ)
1312 so->so_rcv.sb_flags |= SB_AIO;
1313 else
1314 so->so_snd.sb_flags |= SB_AIO;
1315 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1316 ki->kaio_queue_count++;
1317 num_queue_count++;
1318 splx(s);
1319 error = 0;
1320 goto done;
1321 }
1322 splx(s);
1323 }
1324
1325 if ((error = aio_qphysio(p, aiocbe)) == 0)
1326 goto done;
1327 if (error > 0) {
1328 suword(&job->_aiocb_private.status, 0);
1329 aiocbe->uaiocb._aiocb_private.error = error;
1330 suword(&job->_aiocb_private.error, error);
1331 goto done;
1332 }
1333
1334 /* No buffer for daemon I/O. */
1335 aiocbe->bp = NULL;
1336
1337 ki->kaio_queue_count++;
1338 if (lj)
1339 lj->lioj_queue_count++;
1340 s = splnet();
1341 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1342 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1343 splx(s);
1344 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1345
1346 num_queue_count++;
1347 error = 0;
1348
1349 /*
1350 * If we don't have a free AIO process, and we are below our quota, then
1351 * start one. Otherwise, depend on the subsequent I/O completions to
d0d91865 1352 * pick-up this job. If we don't successfully create the new process
984263bc
MD
1353 * (thread) due to resource issues, we return an error for now (EAGAIN),
1354 * which is likely not the correct thing to do.
1355 */
1356 s = splnet();
1357retryproc:
1358 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1359 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1360 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1361 aiop->aioprocflags &= ~AIOP_FREE;
1362 wakeup(aiop->aioproc);
1363 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1364 ((ki->kaio_active_count + num_aio_resv_start) <
1365 ki->kaio_maxactive_count)) {
1366 num_aio_resv_start++;
1367 if ((error = aio_newproc()) == 0) {
1368 num_aio_resv_start--;
1369 goto retryproc;
1370 }
1371 num_aio_resv_start--;
1372 }
1373 splx(s);
1374done:
1375 return error;
1376}
1377
1378/*
1379 * This routine queues an AIO request, checking for quotas.
1380 */
1381static int
41c20dac 1382aio_aqueue(struct aiocb *job, int type)
984263bc 1383{
7b95be2a 1384 struct proc *p = curproc;
984263bc
MD
1385 struct kaioinfo *ki;
1386
1387 if (p->p_aioinfo == NULL)
1388 aio_init_aioinfo(p);
1389
1390 if (num_queue_count >= max_queue_count)
1391 return EAGAIN;
1392
1393 ki = p->p_aioinfo;
1394 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1395 return EAGAIN;
1396
41c20dac 1397 return _aio_aqueue(job, NULL, type);
984263bc
MD
1398}
1399#endif /* VFS_AIO */
1400
1401/*
1402 * Support the aio_return system call, as a side-effect, kernel resources are
1403 * released.
1404 */
1405int
41c20dac 1406aio_return(struct aio_return_args *uap)
984263bc
MD
1407{
1408#ifndef VFS_AIO
1409 return ENOSYS;
1410#else
41c20dac 1411 struct proc *p = curproc;
984263bc
MD
1412 int s;
1413 long jobref;
1414 struct aiocblist *cb, *ncb;
1415 struct aiocb *ujob;
1416 struct kaioinfo *ki;
1417
1418 ki = p->p_aioinfo;
1419 if (ki == NULL)
1420 return EINVAL;
1421
1422 ujob = uap->aiocbp;
1423
1424 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1425 if (jobref == -1 || jobref == 0)
1426 return EINVAL;
1427
1428 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1429 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1430 jobref) {
1431 if (ujob == cb->uuaiocb) {
c7114eea 1432 uap->sysmsg_result =
984263bc
MD
1433 cb->uaiocb._aiocb_private.status;
1434 } else
c7114eea 1435 uap->sysmsg_result = EFAULT;
984263bc
MD
1436 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1437 p->p_stats->p_ru.ru_oublock +=
1438 cb->outputcharge;
1439 cb->outputcharge = 0;
1440 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1441 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1442 cb->inputcharge = 0;
1443 }
1444 aio_free_entry(cb);
1445 return 0;
1446 }
1447 }
1448 s = splbio();
1449 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1450 ncb = TAILQ_NEXT(cb, plist);
1451 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1452 == jobref) {
1453 splx(s);
1454 if (ujob == cb->uuaiocb) {
c7114eea 1455 uap->sysmsg_result =
984263bc
MD
1456 cb->uaiocb._aiocb_private.status;
1457 } else
c7114eea 1458 uap->sysmsg_result = EFAULT;
984263bc
MD
1459 aio_free_entry(cb);
1460 return 0;
1461 }
1462 }
1463 splx(s);
1464
1465 return (EINVAL);
1466#endif /* VFS_AIO */
1467}
1468
1469/*
1470 * Allow a process to wakeup when any of the I/O requests are completed.
1471 */
1472int
41c20dac 1473aio_suspend(struct aio_suspend_args *uap)
984263bc
MD
1474{
1475#ifndef VFS_AIO
1476 return ENOSYS;
1477#else
41c20dac 1478 struct proc *p = curproc;
984263bc
MD
1479 struct timeval atv;
1480 struct timespec ts;
1481 struct aiocb *const *cbptr, *cbp;
1482 struct kaioinfo *ki;
1483 struct aiocblist *cb;
1484 int i;
1485 int njoblist;
1486 int error, s, timo;
1487 long *ijoblist;
1488 struct aiocb **ujoblist;
1489
1490 if (uap->nent > AIO_LISTIO_MAX)
1491 return EINVAL;
1492
1493 timo = 0;
1494 if (uap->timeout) {
1495 /* Get timespec struct. */
1496 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1497 return error;
1498
1499 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1500 return (EINVAL);
1501
1502 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1503 if (itimerfix(&atv))
1504 return (EINVAL);
a94976ad 1505 timo = tvtohz_high(&atv);
984263bc
MD
1506 }
1507
1508 ki = p->p_aioinfo;
1509 if (ki == NULL)
1510 return EAGAIN;
1511
1512 njoblist = 0;
1513 ijoblist = zalloc(aiol_zone);
1514 ujoblist = zalloc(aiol_zone);
1515 cbptr = uap->aiocbp;
1516
1517 for (i = 0; i < uap->nent; i++) {
1518 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1519 if (cbp == 0)
1520 continue;
1521 ujoblist[njoblist] = cbp;
1522 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1523 njoblist++;
1524 }
1525
1526 if (njoblist == 0) {
1527 zfree(aiol_zone, ijoblist);
1528 zfree(aiol_zone, ujoblist);
1529 return 0;
1530 }
1531
1532 error = 0;
1533 for (;;) {
1534 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1535 for (i = 0; i < njoblist; i++) {
1536 if (((intptr_t)
1537 cb->uaiocb._aiocb_private.kernelinfo) ==
1538 ijoblist[i]) {
1539 if (ujoblist[i] != cb->uuaiocb)
1540 error = EINVAL;
1541 zfree(aiol_zone, ijoblist);
1542 zfree(aiol_zone, ujoblist);
1543 return error;
1544 }
1545 }
1546 }
1547
1548 s = splbio();
1549 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1550 TAILQ_NEXT(cb, plist)) {
1551 for (i = 0; i < njoblist; i++) {
1552 if (((intptr_t)
1553 cb->uaiocb._aiocb_private.kernelinfo) ==
1554 ijoblist[i]) {
1555 splx(s);
1556 if (ujoblist[i] != cb->uuaiocb)
1557 error = EINVAL;
1558 zfree(aiol_zone, ijoblist);
1559 zfree(aiol_zone, ujoblist);
1560 return error;
1561 }
1562 }
1563 }
1564
1565 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1566 error = tsleep(p, PCATCH, "aiospn", timo);
984263bc
MD
1567 splx(s);
1568
1569 if (error == ERESTART || error == EINTR) {
1570 zfree(aiol_zone, ijoblist);
1571 zfree(aiol_zone, ujoblist);
1572 return EINTR;
1573 } else if (error == EWOULDBLOCK) {
1574 zfree(aiol_zone, ijoblist);
1575 zfree(aiol_zone, ujoblist);
1576 return EAGAIN;
1577 }
1578 }
1579
1580/* NOTREACHED */
1581 return EINVAL;
1582#endif /* VFS_AIO */
1583}
1584
1585/*
1586 * aio_cancel cancels any non-physio aio operations not currently in
1587 * progress.
1588 */
1589int
41c20dac 1590aio_cancel(struct aio_cancel_args *uap)
984263bc
MD
1591{
1592#ifndef VFS_AIO
1593 return ENOSYS;
1594#else
41c20dac 1595 struct proc *p = curproc;
984263bc
MD
1596 struct kaioinfo *ki;
1597 struct aiocblist *cbe, *cbn;
1598 struct file *fp;
1599 struct filedesc *fdp;
1600 struct socket *so;
1601 struct proc *po;
1602 int s,error;
1603 int cancelled=0;
1604 int notcancelled=0;
1605 struct vnode *vp;
1606
1607 fdp = p->p_fd;
1608 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1609 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1610 return (EBADF);
1611
1612 if (fp->f_type == DTYPE_VNODE) {
1613 vp = (struct vnode *)fp->f_data;
1614
1615 if (vn_isdisk(vp,&error)) {
c7114eea 1616 uap->sysmsg_result = AIO_NOTCANCELED;
984263bc
MD
1617 return 0;
1618 }
1619 } else if (fp->f_type == DTYPE_SOCKET) {
1620 so = (struct socket *)fp->f_data;
1621
1622 s = splnet();
1623
1624 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1625 cbn = TAILQ_NEXT(cbe, list);
1626 if ((uap->aiocbp == NULL) ||
1627 (uap->aiocbp == cbe->uuaiocb) ) {
1628 po = cbe->userproc;
1629 ki = po->p_aioinfo;
1630 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1631 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1632 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1633 if (ki->kaio_flags & KAIO_WAKEUP) {
1634 wakeup(po);
1635 }
1636 cbe->jobstate = JOBST_JOBFINISHED;
1637 cbe->uaiocb._aiocb_private.status=-1;
1638 cbe->uaiocb._aiocb_private.error=ECANCELED;
1639 cancelled++;
1640/* XXX cancelled, knote? */
1641 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1642 SIGEV_SIGNAL)
1643 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1644 if (uap->aiocbp)
1645 break;
1646 }
1647 }
1648 splx(s);
1649
1650 if ((cancelled) && (uap->aiocbp)) {
c7114eea 1651 uap->sysmsg_result = AIO_CANCELED;
984263bc
MD
1652 return 0;
1653 }
1654 }
1655 ki=p->p_aioinfo;
1656 if (ki == NULL)
1657 goto done;
1658 s = splnet();
1659
1660 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1661 cbn = TAILQ_NEXT(cbe, plist);
1662
1663 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1664 ((uap->aiocbp == NULL ) ||
1665 (uap->aiocbp == cbe->uuaiocb))) {
1666
1667 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1668 TAILQ_REMOVE(&aio_jobs, cbe, list);
1669 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1670 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1671 plist);
1672 cancelled++;
1673 ki->kaio_queue_finished_count++;
1674 cbe->jobstate = JOBST_JOBFINISHED;
1675 cbe->uaiocb._aiocb_private.status = -1;
1676 cbe->uaiocb._aiocb_private.error = ECANCELED;
1677/* XXX cancelled, knote? */
1678 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1679 SIGEV_SIGNAL)
1680 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1681 } else {
1682 notcancelled++;
1683 }
1684 }
1685 }
1686 splx(s);
1687done:
1688 if (notcancelled) {
c7114eea 1689 uap->sysmsg_result = AIO_NOTCANCELED;
984263bc
MD
1690 return 0;
1691 }
1692 if (cancelled) {
c7114eea 1693 uap->sysmsg_result = AIO_CANCELED;
984263bc
MD
1694 return 0;
1695 }
c7114eea 1696 uap->sysmsg_result = AIO_ALLDONE;
984263bc
MD
1697
1698 return 0;
1699#endif /* VFS_AIO */
1700}
1701
1702/*
1703 * aio_error is implemented in the kernel level for compatibility purposes only.
1704 * For a user mode async implementation, it would be best to do it in a userland
1705 * subroutine.
1706 */
1707int
41c20dac 1708aio_error(struct aio_error_args *uap)
984263bc
MD
1709{
1710#ifndef VFS_AIO
1711 return ENOSYS;
1712#else
41c20dac 1713 struct proc *p = curproc;
984263bc
MD
1714 int s;
1715 struct aiocblist *cb;
1716 struct kaioinfo *ki;
1717 long jobref;
1718
1719 ki = p->p_aioinfo;
1720 if (ki == NULL)
1721 return EINVAL;
1722
1723 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1724 if ((jobref == -1) || (jobref == 0))
1725 return EINVAL;
1726
1727 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1728 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1729 jobref) {
c7114eea 1730 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
984263bc
MD
1731 return 0;
1732 }
1733 }
1734
1735 s = splnet();
1736
1737 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1738 plist)) {
1739 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1740 jobref) {
c7114eea 1741 uap->sysmsg_result = EINPROGRESS;
984263bc
MD
1742 splx(s);
1743 return 0;
1744 }
1745 }
1746
1747 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1748 plist)) {
1749 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1750 jobref) {
c7114eea 1751 uap->sysmsg_result = EINPROGRESS;
984263bc
MD
1752 splx(s);
1753 return 0;
1754 }
1755 }
1756 splx(s);
1757
1758 s = splbio();
1759 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1760 plist)) {
1761 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1762 jobref) {
c7114eea 1763 uap->sysmsg_result = cb->uaiocb._aiocb_private.error;
984263bc
MD
1764 splx(s);
1765 return 0;
1766 }
1767 }
1768
1769 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1770 plist)) {
1771 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1772 jobref) {
c7114eea 1773 uap->sysmsg_result = EINPROGRESS;
984263bc
MD
1774 splx(s);
1775 return 0;
1776 }
1777 }
1778 splx(s);
1779
1780#if (0)
1781 /*
1782 * Hack for lio.
1783 */
1784 status = fuword(&uap->aiocbp->_aiocb_private.status);
1785 if (status == -1)
1786 return fuword(&uap->aiocbp->_aiocb_private.error);
1787#endif
1788 return EINVAL;
1789#endif /* VFS_AIO */
1790}
1791
1792/* syscall - asynchronous read from a file (REALTIME) */
1793int
41c20dac 1794aio_read(struct aio_read_args *uap)
984263bc
MD
1795{
1796#ifndef VFS_AIO
1797 return ENOSYS;
1798#else
41c20dac 1799 return aio_aqueue(uap->aiocbp, LIO_READ);
984263bc
MD
1800#endif /* VFS_AIO */
1801}
1802
1803/* syscall - asynchronous write to a file (REALTIME) */
1804int
41c20dac 1805aio_write(struct aio_write_args *uap)
984263bc
MD
1806{
1807#ifndef VFS_AIO
1808 return ENOSYS;
1809#else
41c20dac 1810 return aio_aqueue(uap->aiocbp, LIO_WRITE);
984263bc
MD
1811#endif /* VFS_AIO */
1812}
1813
1814/* syscall - XXX undocumented */
1815int
41c20dac 1816lio_listio(struct lio_listio_args *uap)
984263bc
MD
1817{
1818#ifndef VFS_AIO
1819 return ENOSYS;
1820#else
41c20dac 1821 struct proc *p = curproc;
984263bc
MD
1822 int nent, nentqueued;
1823 struct aiocb *iocb, * const *cbptr;
1824 struct aiocblist *cb;
1825 struct kaioinfo *ki;
1826 struct aio_liojob *lj;
1827 int error, runningcode;
1828 int nerror;
1829 int i;
1830 int s;
1831
1832 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1833 return EINVAL;
1834
1835 nent = uap->nent;
1836 if (nent > AIO_LISTIO_MAX)
1837 return EINVAL;
1838
1839 if (p->p_aioinfo == NULL)
1840 aio_init_aioinfo(p);
1841
1842 if ((nent + num_queue_count) > max_queue_count)
1843 return EAGAIN;
1844
1845 ki = p->p_aioinfo;
1846 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1847 return EAGAIN;
1848
1849 lj = zalloc(aiolio_zone);
1850 if (!lj)
1851 return EAGAIN;
1852
1853 lj->lioj_flags = 0;
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;
1858 lj->lioj_ki = ki;
1859
1860 /*
1861 * Setup signal.
1862 */
1863 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1864 error = copyin(uap->sig, &lj->lioj_signal,
1865 sizeof(lj->lioj_signal));
1866 if (error) {
1867 zfree(aiolio_zone, lj);
1868 return error;
1869 }
1870 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1871 zfree(aiolio_zone, lj);
1872 return EINVAL;
1873 }
1874 lj->lioj_flags |= LIOJ_SIGNAL;
1875 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1876 } else
1877 lj->lioj_flags &= ~LIOJ_SIGNAL;
1878
1879 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1880 /*
1881 * Get pointers to the list of I/O requests.
1882 */
1883 nerror = 0;
1884 nentqueued = 0;
1885 cbptr = uap->acb_list;
1886 for (i = 0; i < uap->nent; i++) {
1887 iocb = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1888 if (((intptr_t)iocb != -1) && ((intptr_t)iocb != 0)) {
41c20dac 1889 error = _aio_aqueue(iocb, lj, 0);
984263bc
MD
1890 if (error == 0)
1891 nentqueued++;
1892 else
1893 nerror++;
1894 }
1895 }
1896
1897 /*
1898 * If we haven't queued any, then just return error.
1899 */
1900 if (nentqueued == 0)
1901 return 0;
1902
1903 /*
1904 * Calculate the appropriate error return.
1905 */
1906 runningcode = 0;
1907 if (nerror)
1908 runningcode = EIO;
1909
1910 if (uap->mode == LIO_WAIT) {
1911 int command, found, jobref;
1912
1913 for (;;) {
1914 found = 0;
1915 for (i = 0; i < uap->nent; i++) {
1916 /*
1917 * Fetch address of the control buf pointer in
1918 * user space.
1919 */
1920 iocb = (struct aiocb *)
1921 (intptr_t)fuword(&cbptr[i]);
1922 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1923 == 0))
1924 continue;
1925
1926 /*
1927 * Fetch the associated command from user space.
1928 */
1929 command = fuword(&iocb->aio_lio_opcode);
1930 if (command == LIO_NOP) {
1931 found++;
1932 continue;
1933 }
1934
1935 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1936
1937 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1938 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1939 == jobref) {
1940 if (cb->uaiocb.aio_lio_opcode
1941 == LIO_WRITE) {
1942 p->p_stats->p_ru.ru_oublock
1943 +=
1944 cb->outputcharge;
1945 cb->outputcharge = 0;
1946 } else if (cb->uaiocb.aio_lio_opcode
1947 == LIO_READ) {
1948 p->p_stats->p_ru.ru_inblock
1949 += cb->inputcharge;
1950 cb->inputcharge = 0;
1951 }
1952 found++;
1953 break;
1954 }
1955 }
1956
1957 s = splbio();
1958 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1959 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1960 == jobref) {
1961 found++;
1962 break;
1963 }
1964 }
1965 splx(s);
1966 }
1967
1968 /*
1969 * If all I/Os have been disposed of, then we can
1970 * return.
1971 */
1972 if (found == nentqueued)
1973 return runningcode;
1974
1975 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1976 error = tsleep(p, PCATCH, "aiospn", 0);
984263bc
MD
1977
1978 if (error == EINTR)
1979 return EINTR;
1980 else if (error == EWOULDBLOCK)
1981 return EAGAIN;
1982 }
1983 }
1984
1985 return runningcode;
1986#endif /* VFS_AIO */
1987}
1988
1989#ifdef VFS_AIO
1990/*
1991 * This is a weird hack so that we can post a signal. It is safe to do so from
1992 * a timeout routine, but *not* from an interrupt routine.
1993 */
1994static void
1995process_signal(void *aioj)
1996{
1997 struct aiocblist *aiocbe = aioj;
1998 struct aio_liojob *lj = aiocbe->lio;
1999 struct aiocb *cb = &aiocbe->uaiocb;
2000
2001 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2002 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2003 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2004 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2005 }
2006
2007 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2008 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2009}
2010
2011/*
2012 * Interrupt handler for physio, performs the necessary process wakeups, and
2013 * signals.
2014 */
2015static void
2016aio_physwakeup(struct buf *bp)
2017{
2018 struct aiocblist *aiocbe;
2019 struct proc *p;
2020 struct kaioinfo *ki;
2021 struct aio_liojob *lj;
2022
2023 wakeup(bp);
2024
2025 aiocbe = (struct aiocblist *)bp->b_spc;
2026 if (aiocbe) {
2027 p = bp->b_caller1;
2028
2029 aiocbe->jobstate = JOBST_JOBBFINISHED;
2030 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2031 aiocbe->uaiocb._aiocb_private.error = 0;
2032 aiocbe->jobflags |= AIOCBLIST_DONE;
2033
2034 if (bp->b_flags & B_ERROR)
2035 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2036
2037 lj = aiocbe->lio;
2038 if (lj) {
2039 lj->lioj_buffer_finished_count++;
2040
2041 /*
2042 * wakeup/signal if all of the interrupt jobs are done.
2043 */
2044 if (lj->lioj_buffer_finished_count ==
2045 lj->lioj_buffer_count) {
2046 /*
2047 * Post a signal if it is called for.
2048 */
2049 if ((lj->lioj_flags &
2050 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2051 LIOJ_SIGNAL) {
2052 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
a5eb27b6
MD
2053 callout_reset(&aiocbe->timeout, 0,
2054 process_signal, aiocbe);
984263bc
MD
2055 }
2056 }
2057 }
2058
2059 ki = p->p_aioinfo;
2060 if (ki) {
2061 ki->kaio_buffer_finished_count++;
2062 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2063 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2064 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2065
2066 KNOTE(&aiocbe->klist, 0);
2067 /* Do the wakeup. */
2068 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2069 ki->kaio_flags &= ~KAIO_WAKEUP;
2070 wakeup(p);
2071 }
2072 }
2073
a5eb27b6
MD
2074 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
2075 callout_reset(&aiocbe->timeout, 0,
2076 process_signal, aiocbe);
2077 }
984263bc
MD
2078 }
2079}
2080#endif /* VFS_AIO */
2081
2082/* syscall - wait for the next completion of an aio request */
2083int
41c20dac 2084aio_waitcomplete(struct aio_waitcomplete_args *uap)
984263bc
MD
2085{
2086#ifndef VFS_AIO
2087 return ENOSYS;
2088#else
41c20dac 2089 struct proc *p = curproc;
984263bc
MD
2090 struct timeval atv;
2091 struct timespec ts;
2092 struct kaioinfo *ki;
2093 struct aiocblist *cb = NULL;
2094 int error, s, timo;
2095
2096 suword(uap->aiocbp, (int)NULL);
2097
2098 timo = 0;
2099 if (uap->timeout) {
2100 /* Get timespec struct. */
2101 error = copyin(uap->timeout, &ts, sizeof(ts));
2102 if (error)
2103 return error;
2104
2105 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2106 return (EINVAL);
2107
2108 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2109 if (itimerfix(&atv))
2110 return (EINVAL);
a94976ad 2111 timo = tvtohz_high(&atv);
984263bc
MD
2112 }
2113
2114 ki = p->p_aioinfo;
2115 if (ki == NULL)
2116 return EAGAIN;
2117
2118 for (;;) {
2119 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2120 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
c7114eea 2121 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2122 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2123 p->p_stats->p_ru.ru_oublock +=
2124 cb->outputcharge;
2125 cb->outputcharge = 0;
2126 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2127 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2128 cb->inputcharge = 0;
2129 }
2130 aio_free_entry(cb);
2131 return cb->uaiocb._aiocb_private.error;
2132 }
2133
2134 s = splbio();
2135 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2136 splx(s);
2137 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
c7114eea 2138 uap->sysmsg_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2139 aio_free_entry(cb);
2140 return cb->uaiocb._aiocb_private.error;
2141 }
2142
2143 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 2144 error = tsleep(p, PCATCH, "aiowc", timo);
984263bc
MD
2145 splx(s);
2146
2147 if (error == ERESTART)
2148 return EINTR;
2149 else if (error < 0)
2150 return error;
2151 else if (error == EINTR)
2152 return EINTR;
2153 else if (error == EWOULDBLOCK)
2154 return EAGAIN;
2155 }
2156#endif /* VFS_AIO */
2157}
2158
2159#ifndef VFS_AIO
2160static int
2161filt_aioattach(struct knote *kn)
2162{
2163
2164 return (ENXIO);
2165}
2166
2167struct filterops aio_filtops =
2168 { 0, filt_aioattach, NULL, NULL };
2169
2170#else
2171/* kqueue attach function */
2172static int
2173filt_aioattach(struct knote *kn)
2174{
2175 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2176
2177 /*
2178 * The aiocbe pointer must be validated before using it, so
2179 * registration is restricted to the kernel; the user cannot
2180 * set EV_FLAG1.
2181 */
2182 if ((kn->kn_flags & EV_FLAG1) == 0)
2183 return (EPERM);
2184 kn->kn_flags &= ~EV_FLAG1;
2185
2186 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2187
2188 return (0);
2189}
2190
2191/* kqueue detach function */
2192static void
2193filt_aiodetach(struct knote *kn)
2194{
2195 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2196
2197 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2198}
2199
2200/* kqueue filter function */
2201/*ARGSUSED*/
2202static int
2203filt_aio(struct knote *kn, long hint)
2204{
2205 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2206
2207 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2208 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2209 aiocbe->jobstate != JOBST_JOBBFINISHED)
2210 return (0);
2211 kn->kn_flags |= EV_EOF;
2212 return (1);
2213}
2214
2215struct filterops aio_filtops =
2216 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2217#endif /* VFS_AIO */