Fix CPU stats percentages formatting
[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 $
90b9818c 17 * $DragonFly: src/sys/kern/vfs_aio.c,v 1.9 2003/07/26 18:12:44 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;
388 untimeout(process_signal, aiocbe, aiocbe->timeouthandle);
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;
673 uifree(cr->cr_uidinfo);
674 cr->cr_uidinfo = uifind(0);
675 cr->cr_ngroups = 1;
676 cr->cr_groups[0] = 1;
984263bc
MD
677
678 /* The daemon resides in its own pgrp. */
679 enterpgrp(mycp, mycp->p_pid, 1);
680
681 /* Mark special process type. */
682 mycp->p_flag |= P_SYSTEM | P_KTHREADP;
683
684 /*
685 * Wakeup parent process. (Parent sleeps to keep from blasting away
686 * and creating too many daemons.)
687 */
688 wakeup(mycp);
689
690 for (;;) {
691 /*
692 * curcp is the current daemon process context.
693 * userp is the current user process context.
694 */
695 curcp = mycp;
696
697 /*
698 * Take daemon off of free queue
699 */
700 if (aiop->aioprocflags & AIOP_FREE) {
701 s = splnet();
702 TAILQ_REMOVE(&aio_freeproc, aiop, list);
703 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
704 aiop->aioprocflags &= ~AIOP_FREE;
705 splx(s);
706 }
707 aiop->aioprocflags &= ~AIOP_SCHED;
708
709 /*
710 * Check for jobs.
711 */
712 while ((aiocbe = aio_selectjob(aiop)) != NULL) {
713 cb = &aiocbe->uaiocb;
714 userp = aiocbe->userproc;
715
716 aiocbe->jobstate = JOBST_JOBRUNNING;
717
718 /*
719 * Connect to process address space for user program.
720 */
721 if (userp != curcp) {
722 /*
723 * Save the current address space that we are
724 * connected to.
725 */
726 tmpvm = mycp->p_vmspace;
727
728 /*
729 * Point to the new user address space, and
730 * refer to it.
731 */
732 mycp->p_vmspace = userp->p_vmspace;
733 mycp->p_vmspace->vm_refcnt++;
734
735 /* Activate the new mapping. */
736 pmap_activate(mycp);
737
738 /*
739 * If the old address space wasn't the daemons
740 * own address space, then we need to remove the
741 * daemon's reference from the other process
742 * that it was acting on behalf of.
743 */
744 if (tmpvm != myvm) {
745 vmspace_free(tmpvm);
746 }
747 curcp = userp;
748 }
749
750 ki = userp->p_aioinfo;
751 lj = aiocbe->lio;
752
753 /* Account for currently active jobs. */
754 ki->kaio_active_count++;
755
756 /* Do the I/O function. */
757 aio_process(aiocbe);
758
759 /* Decrement the active job count. */
760 ki->kaio_active_count--;
761
762 /*
763 * Increment the completion count for wakeup/signal
764 * comparisons.
765 */
766 aiocbe->jobflags |= AIOCBLIST_DONE;
767 ki->kaio_queue_finished_count++;
768 if (lj)
769 lj->lioj_queue_finished_count++;
770 if ((ki->kaio_flags & KAIO_WAKEUP) || ((ki->kaio_flags
771 & KAIO_RUNDOWN) && (ki->kaio_active_count == 0))) {
772 ki->kaio_flags &= ~KAIO_WAKEUP;
773 wakeup(userp);
774 }
775
776 s = splbio();
777 if (lj && (lj->lioj_flags &
778 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) == LIOJ_SIGNAL) {
779 if ((lj->lioj_queue_finished_count ==
780 lj->lioj_queue_count) &&
781 (lj->lioj_buffer_finished_count ==
782 lj->lioj_buffer_count)) {
783 psignal(userp,
784 lj->lioj_signal.sigev_signo);
785 lj->lioj_flags |=
786 LIOJ_SIGNAL_POSTED;
787 }
788 }
789 splx(s);
790
791 aiocbe->jobstate = JOBST_JOBFINISHED;
792
793 s = splnet();
794 TAILQ_REMOVE(&ki->kaio_jobqueue, aiocbe, plist);
795 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, aiocbe, plist);
796 splx(s);
797 KNOTE(&aiocbe->klist, 0);
798
799 if (aiocbe->jobflags & AIOCBLIST_RUNDOWN) {
800 wakeup(aiocbe);
801 aiocbe->jobflags &= ~AIOCBLIST_RUNDOWN;
802 }
803
804 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL) {
805 psignal(userp, cb->aio_sigevent.sigev_signo);
806 }
807 }
808
809 /*
810 * Disconnect from user address space.
811 */
812 if (curcp != mycp) {
813 /* Get the user address space to disconnect from. */
814 tmpvm = mycp->p_vmspace;
815
816 /* Get original address space for daemon. */
817 mycp->p_vmspace = myvm;
818
819 /* Activate the daemon's address space. */
820 pmap_activate(mycp);
821#ifdef DIAGNOSTIC
822 if (tmpvm == myvm) {
823 printf("AIOD: vmspace problem -- %d\n",
824 mycp->p_pid);
825 }
826#endif
827 /* Remove our vmspace reference. */
828 vmspace_free(tmpvm);
829
830 curcp = mycp;
831 }
832
833 /*
834 * If we are the first to be put onto the free queue, wakeup
835 * anyone waiting for a daemon.
836 */
837 s = splnet();
838 TAILQ_REMOVE(&aio_activeproc, aiop, list);
839 if (TAILQ_EMPTY(&aio_freeproc))
840 wakeup(&aio_freeproc);
841 TAILQ_INSERT_HEAD(&aio_freeproc, aiop, list);
842 aiop->aioprocflags |= AIOP_FREE;
843 splx(s);
844
845 /*
846 * If daemon is inactive for a long time, allow it to exit,
847 * thereby freeing resources.
848 */
849 if (((aiop->aioprocflags & AIOP_SCHED) == 0) && tsleep(mycp,
377d4740 850 0, "aiordy", aiod_lifetime)) {
984263bc
MD
851 s = splnet();
852 if (TAILQ_EMPTY(&aio_jobs)) {
853 if ((aiop->aioprocflags & AIOP_FREE) &&
854 (num_aio_procs > target_aio_procs)) {
855 TAILQ_REMOVE(&aio_freeproc, aiop, list);
856 splx(s);
857 zfree(aiop_zone, aiop);
858 num_aio_procs--;
859#ifdef DIAGNOSTIC
860 if (mycp->p_vmspace->vm_refcnt <= 1) {
861 printf("AIOD: bad vm refcnt for"
862 " exiting daemon: %d\n",
863 mycp->p_vmspace->vm_refcnt);
864 }
865#endif
7b95be2a 866 exit1(0);
984263bc
MD
867 }
868 }
869 splx(s);
870 }
871 }
872}
873
874/*
875 * Create a new AIO daemon. This is mostly a kernel-thread fork routine. The
876 * AIO daemon modifies its environment itself.
877 */
878static int
879aio_newproc()
880{
881 int error;
882 struct proc *p, *np;
883
884 p = &proc0;
885 error = fork1(p, RFPROC|RFMEM|RFNOWAIT, &np);
886 if (error)
887 return error;
888 cpu_set_fork_handler(np, aio_daemon, curproc);
7d0bac62 889 start_forked_proc(p, np);
984263bc
MD
890
891 /*
892 * Wait until daemon is started, but continue on just in case to
893 * handle error conditions.
894 */
377d4740 895 error = tsleep(np, 0, "aiosta", aiod_timeout);
984263bc
MD
896 num_aio_procs++;
897
898 return error;
899}
900
901/*
902 * Try the high-performance, low-overhead physio method for eligible
903 * VCHR devices. This method doesn't use an aio helper thread, and
904 * thus has very low overhead.
905 *
906 * Assumes that the caller, _aio_aqueue(), has incremented the file
907 * structure's reference count, preventing its deallocation for the
908 * duration of this call.
909 */
910static int
911aio_qphysio(struct proc *p, struct aiocblist *aiocbe)
912{
913 int error;
914 struct aiocb *cb;
915 struct file *fp;
916 struct buf *bp;
917 struct vnode *vp;
918 struct kaioinfo *ki;
919 struct aio_liojob *lj;
920 int s;
921 int notify;
922
923 cb = &aiocbe->uaiocb;
924 fp = aiocbe->fd_file;
925
926 if (fp->f_type != DTYPE_VNODE)
927 return (-1);
928
929 vp = (struct vnode *)fp->f_data;
930
931 /*
932 * If its not a disk, we don't want to return a positive error.
933 * It causes the aio code to not fall through to try the thread
934 * way when you're talking to a regular file.
935 */
936 if (!vn_isdisk(vp, &error)) {
937 if (error == ENOTBLK)
938 return (-1);
939 else
940 return (error);
941 }
942
943 if (cb->aio_nbytes % vp->v_rdev->si_bsize_phys)
944 return (-1);
945
946 if (cb->aio_nbytes >
947 MAXPHYS - (((vm_offset_t) cb->aio_buf) & PAGE_MASK))
948 return (-1);
949
950 ki = p->p_aioinfo;
951 if (ki->kaio_buffer_count >= ki->kaio_ballowed_count)
952 return (-1);
953
954 ki->kaio_buffer_count++;
955
956 lj = aiocbe->lio;
957 if (lj)
958 lj->lioj_buffer_count++;
959
960 /* Create and build a buffer header for a transfer. */
961 bp = (struct buf *)getpbuf(NULL);
962 BUF_KERNPROC(bp);
963
964 /*
965 * Get a copy of the kva from the physical buffer.
966 */
967 bp->b_caller1 = p;
968 bp->b_dev = vp->v_rdev;
969 error = 0;
970
971 bp->b_bcount = cb->aio_nbytes;
972 bp->b_bufsize = cb->aio_nbytes;
973 bp->b_flags = B_PHYS | B_CALL | (cb->aio_lio_opcode == LIO_WRITE ?
974 B_WRITE : B_READ);
975 bp->b_iodone = aio_physwakeup;
976 bp->b_saveaddr = bp->b_data;
977 bp->b_data = (void *)(uintptr_t)cb->aio_buf;
978 bp->b_blkno = btodb(cb->aio_offset);
979
980 /* Bring buffer into kernel space. */
981 if (vmapbuf(bp) < 0) {
982 error = EFAULT;
983 goto doerror;
984 }
985
986 s = splbio();
987 aiocbe->bp = bp;
988 bp->b_spc = (void *)aiocbe;
989 TAILQ_INSERT_TAIL(&aio_bufjobs, aiocbe, list);
990 TAILQ_INSERT_TAIL(&ki->kaio_bufqueue, aiocbe, plist);
991 aiocbe->jobstate = JOBST_JOBQBUF;
992 cb->_aiocb_private.status = cb->aio_nbytes;
993 num_buf_aio++;
994 bp->b_error = 0;
995
996 splx(s);
997
998 /* Perform transfer. */
999 BUF_STRATEGY(bp, 0);
1000
1001 notify = 0;
1002 s = splbio();
1003
1004 /*
1005 * If we had an error invoking the request, or an error in processing
1006 * the request before we have returned, we process it as an error in
1007 * transfer. Note that such an I/O error is not indicated immediately,
1008 * but is returned using the aio_error mechanism. In this case,
1009 * aio_suspend will return immediately.
1010 */
1011 if (bp->b_error || (bp->b_flags & B_ERROR)) {
1012 struct aiocb *job = aiocbe->uuaiocb;
1013
1014 aiocbe->uaiocb._aiocb_private.status = 0;
1015 suword(&job->_aiocb_private.status, 0);
1016 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
1017 suword(&job->_aiocb_private.error, bp->b_error);
1018
1019 ki->kaio_buffer_finished_count++;
1020
1021 if (aiocbe->jobstate != JOBST_JOBBFINISHED) {
1022 aiocbe->jobstate = JOBST_JOBBFINISHED;
1023 aiocbe->jobflags |= AIOCBLIST_DONE;
1024 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
1025 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
1026 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
1027 notify = 1;
1028 }
1029 }
1030 splx(s);
1031 if (notify)
1032 KNOTE(&aiocbe->klist, 0);
1033 return 0;
1034
1035doerror:
1036 ki->kaio_buffer_count--;
1037 if (lj)
1038 lj->lioj_buffer_count--;
1039 aiocbe->bp = NULL;
1040 relpbuf(bp, NULL);
1041 return error;
1042}
1043
1044/*
1045 * This waits/tests physio completion.
1046 */
1047static int
1048aio_fphysio(struct aiocblist *iocb)
1049{
1050 int s;
1051 struct buf *bp;
1052 int error;
1053
1054 bp = iocb->bp;
1055
1056 s = splbio();
1057 while ((bp->b_flags & B_DONE) == 0) {
377d4740 1058 if (tsleep(bp, 0, "physstr", aiod_timeout)) {
984263bc
MD
1059 if ((bp->b_flags & B_DONE) == 0) {
1060 splx(s);
1061 return EINPROGRESS;
1062 } else
1063 break;
1064 }
1065 }
1066 splx(s);
1067
1068 /* Release mapping into kernel space. */
1069 vunmapbuf(bp);
1070 iocb->bp = 0;
1071
1072 error = 0;
1073
1074 /* Check for an error. */
1075 if (bp->b_flags & B_ERROR)
1076 error = bp->b_error;
1077
1078 relpbuf(bp, NULL);
1079 return (error);
1080}
1081#endif /* VFS_AIO */
1082
1083/*
1084 * Wake up aio requests that may be serviceable now.
1085 */
1086void
1087aio_swake(struct socket *so, struct sockbuf *sb)
1088{
1089#ifndef VFS_AIO
1090 return;
1091#else
1092 struct aiocblist *cb,*cbn;
1093 struct proc *p;
1094 struct kaioinfo *ki = NULL;
1095 int opcode, wakecount = 0;
1096 struct aioproclist *aiop;
1097
1098 if (sb == &so->so_snd) {
1099 opcode = LIO_WRITE;
1100 so->so_snd.sb_flags &= ~SB_AIO;
1101 } else {
1102 opcode = LIO_READ;
1103 so->so_rcv.sb_flags &= ~SB_AIO;
1104 }
1105
1106 for (cb = TAILQ_FIRST(&so->so_aiojobq); cb; cb = cbn) {
1107 cbn = TAILQ_NEXT(cb, list);
1108 if (opcode == cb->uaiocb.aio_lio_opcode) {
1109 p = cb->userproc;
1110 ki = p->p_aioinfo;
1111 TAILQ_REMOVE(&so->so_aiojobq, cb, list);
1112 TAILQ_REMOVE(&ki->kaio_sockqueue, cb, plist);
1113 TAILQ_INSERT_TAIL(&aio_jobs, cb, list);
1114 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, cb, plist);
1115 wakecount++;
1116 if (cb->jobstate != JOBST_JOBQGLOBAL)
1117 panic("invalid queue value");
1118 }
1119 }
1120
1121 while (wakecount--) {
1122 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != 0) {
1123 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1124 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1125 aiop->aioprocflags &= ~AIOP_FREE;
1126 wakeup(aiop->aioproc);
1127 }
1128 }
1129#endif /* VFS_AIO */
1130}
1131
1132#ifdef VFS_AIO
1133/*
1134 * Queue a new AIO request. Choosing either the threaded or direct physio VCHR
1135 * technique is done in this code.
1136 */
1137static int
41c20dac 1138_aio_aqueue(struct aiocb *job, struct aio_liojob *lj, int type)
984263bc 1139{
7b95be2a 1140 struct proc *p = curproc;
984263bc
MD
1141 struct filedesc *fdp;
1142 struct file *fp;
1143 unsigned int fd;
1144 struct socket *so;
1145 int s;
1146 int error;
1147 int opcode, user_opcode;
1148 struct aiocblist *aiocbe;
1149 struct aioproclist *aiop;
1150 struct kaioinfo *ki;
1151 struct kevent kev;
1152 struct kqueue *kq;
1153 struct file *kq_fp;
1154
1155 if ((aiocbe = TAILQ_FIRST(&aio_freejobs)) != NULL)
1156 TAILQ_REMOVE(&aio_freejobs, aiocbe, list);
1157 else
1158 aiocbe = zalloc (aiocb_zone);
1159
1160 aiocbe->inputcharge = 0;
1161 aiocbe->outputcharge = 0;
1162 callout_handle_init(&aiocbe->timeouthandle);
1163 SLIST_INIT(&aiocbe->klist);
1164
1165 suword(&job->_aiocb_private.status, -1);
1166 suword(&job->_aiocb_private.error, 0);
1167 suword(&job->_aiocb_private.kernelinfo, -1);
1168
1169 error = copyin(job, &aiocbe->uaiocb, sizeof(aiocbe->uaiocb));
1170 if (error) {
1171 suword(&job->_aiocb_private.error, error);
1172 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1173 return error;
1174 }
1175 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL &&
1176 !_SIG_VALID(aiocbe->uaiocb.aio_sigevent.sigev_signo)) {
1177 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1178 return EINVAL;
1179 }
1180
1181 /* Save userspace address of the job info. */
1182 aiocbe->uuaiocb = job;
1183
1184 /* Get the opcode. */
1185 user_opcode = aiocbe->uaiocb.aio_lio_opcode;
1186 if (type != LIO_NOP)
1187 aiocbe->uaiocb.aio_lio_opcode = type;
1188 opcode = aiocbe->uaiocb.aio_lio_opcode;
1189
1190 /* Get the fd info for process. */
1191 fdp = p->p_fd;
1192
1193 /*
1194 * Range check file descriptor.
1195 */
1196 fd = aiocbe->uaiocb.aio_fildes;
1197 if (fd >= fdp->fd_nfiles) {
1198 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1199 if (type == 0)
1200 suword(&job->_aiocb_private.error, EBADF);
1201 return EBADF;
1202 }
1203
1204 fp = aiocbe->fd_file = fdp->fd_ofiles[fd];
1205 if ((fp == NULL) || ((opcode == LIO_WRITE) && ((fp->f_flag & FWRITE) ==
1206 0))) {
1207 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1208 if (type == 0)
1209 suword(&job->_aiocb_private.error, EBADF);
1210 return EBADF;
1211 }
1212 fhold(fp);
1213
1214 if (aiocbe->uaiocb.aio_offset == -1LL) {
1215 error = EINVAL;
1216 goto aqueue_fail;
1217 }
1218 error = suword(&job->_aiocb_private.kernelinfo, jobrefid);
1219 if (error) {
1220 error = EINVAL;
1221 goto aqueue_fail;
1222 }
1223 aiocbe->uaiocb._aiocb_private.kernelinfo = (void *)(intptr_t)jobrefid;
1224 if (jobrefid == LONG_MAX)
1225 jobrefid = 1;
1226 else
1227 jobrefid++;
1228
1229 if (opcode == LIO_NOP) {
7b95be2a 1230 fdrop(fp, p->p_thread);
984263bc
MD
1231 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1232 if (type == 0) {
1233 suword(&job->_aiocb_private.error, 0);
1234 suword(&job->_aiocb_private.status, 0);
1235 suword(&job->_aiocb_private.kernelinfo, 0);
1236 }
1237 return 0;
1238 }
1239 if ((opcode != LIO_READ) && (opcode != LIO_WRITE)) {
1240 if (type == 0)
1241 suword(&job->_aiocb_private.status, 0);
1242 error = EINVAL;
1243 goto aqueue_fail;
1244 }
1245
1246 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_KEVENT) {
1247 kev.ident = aiocbe->uaiocb.aio_sigevent.sigev_notify_kqueue;
1248 kev.udata = aiocbe->uaiocb.aio_sigevent.sigev_value.sigval_ptr;
1249 }
1250 else {
1251 /*
1252 * This method for requesting kevent-based notification won't
1253 * work on the alpha, since we're passing in a pointer
1254 * via aio_lio_opcode, which is an int. Use the SIGEV_KEVENT-
1255 * based method instead.
1256 */
1257 if (user_opcode == LIO_NOP || user_opcode == LIO_READ ||
1258 user_opcode == LIO_WRITE)
1259 goto no_kqueue;
1260
1261 error = copyin((struct kevent *)(uintptr_t)user_opcode,
1262 &kev, sizeof(kev));
1263 if (error)
1264 goto aqueue_fail;
1265 }
1266 if ((u_int)kev.ident >= fdp->fd_nfiles ||
1267 (kq_fp = fdp->fd_ofiles[kev.ident]) == NULL ||
1268 (kq_fp->f_type != DTYPE_KQUEUE)) {
1269 error = EBADF;
1270 goto aqueue_fail;
1271 }
1272 kq = (struct kqueue *)kq_fp->f_data;
1273 kev.ident = (uintptr_t)aiocbe->uuaiocb;
1274 kev.filter = EVFILT_AIO;
1275 kev.flags = EV_ADD | EV_ENABLE | EV_FLAG1;
1276 kev.data = (intptr_t)aiocbe;
7b95be2a 1277 error = kqueue_register(kq, &kev, p->p_thread);
984263bc
MD
1278aqueue_fail:
1279 if (error) {
7b95be2a 1280 fdrop(fp, p->p_thread);
984263bc
MD
1281 TAILQ_INSERT_HEAD(&aio_freejobs, aiocbe, list);
1282 if (type == 0)
1283 suword(&job->_aiocb_private.error, error);
1284 goto done;
1285 }
1286no_kqueue:
1287
1288 suword(&job->_aiocb_private.error, EINPROGRESS);
1289 aiocbe->uaiocb._aiocb_private.error = EINPROGRESS;
1290 aiocbe->userproc = p;
1291 aiocbe->jobflags = 0;
1292 aiocbe->lio = lj;
1293 ki = p->p_aioinfo;
1294
1295 if (fp->f_type == DTYPE_SOCKET) {
1296 /*
1297 * Alternate queueing for socket ops: Reach down into the
1298 * descriptor to get the socket data. Then check to see if the
1299 * socket is ready to be read or written (based on the requested
1300 * operation).
1301 *
1302 * If it is not ready for io, then queue the aiocbe on the
1303 * socket, and set the flags so we get a call when sbnotify()
1304 * happens.
1305 */
1306 so = (struct socket *)fp->f_data;
1307 s = splnet();
1308 if (((opcode == LIO_READ) && (!soreadable(so))) || ((opcode ==
1309 LIO_WRITE) && (!sowriteable(so)))) {
1310 TAILQ_INSERT_TAIL(&so->so_aiojobq, aiocbe, list);
1311 TAILQ_INSERT_TAIL(&ki->kaio_sockqueue, aiocbe, plist);
1312 if (opcode == LIO_READ)
1313 so->so_rcv.sb_flags |= SB_AIO;
1314 else
1315 so->so_snd.sb_flags |= SB_AIO;
1316 aiocbe->jobstate = JOBST_JOBQGLOBAL; /* XXX */
1317 ki->kaio_queue_count++;
1318 num_queue_count++;
1319 splx(s);
1320 error = 0;
1321 goto done;
1322 }
1323 splx(s);
1324 }
1325
1326 if ((error = aio_qphysio(p, aiocbe)) == 0)
1327 goto done;
1328 if (error > 0) {
1329 suword(&job->_aiocb_private.status, 0);
1330 aiocbe->uaiocb._aiocb_private.error = error;
1331 suword(&job->_aiocb_private.error, error);
1332 goto done;
1333 }
1334
1335 /* No buffer for daemon I/O. */
1336 aiocbe->bp = NULL;
1337
1338 ki->kaio_queue_count++;
1339 if (lj)
1340 lj->lioj_queue_count++;
1341 s = splnet();
1342 TAILQ_INSERT_TAIL(&ki->kaio_jobqueue, aiocbe, plist);
1343 TAILQ_INSERT_TAIL(&aio_jobs, aiocbe, list);
1344 splx(s);
1345 aiocbe->jobstate = JOBST_JOBQGLOBAL;
1346
1347 num_queue_count++;
1348 error = 0;
1349
1350 /*
1351 * If we don't have a free AIO process, and we are below our quota, then
1352 * start one. Otherwise, depend on the subsequent I/O completions to
1353 * pick-up this job. If we don't sucessfully create the new process
1354 * (thread) due to resource issues, we return an error for now (EAGAIN),
1355 * which is likely not the correct thing to do.
1356 */
1357 s = splnet();
1358retryproc:
1359 if ((aiop = TAILQ_FIRST(&aio_freeproc)) != NULL) {
1360 TAILQ_REMOVE(&aio_freeproc, aiop, list);
1361 TAILQ_INSERT_TAIL(&aio_activeproc, aiop, list);
1362 aiop->aioprocflags &= ~AIOP_FREE;
1363 wakeup(aiop->aioproc);
1364 } else if (((num_aio_resv_start + num_aio_procs) < max_aio_procs) &&
1365 ((ki->kaio_active_count + num_aio_resv_start) <
1366 ki->kaio_maxactive_count)) {
1367 num_aio_resv_start++;
1368 if ((error = aio_newproc()) == 0) {
1369 num_aio_resv_start--;
1370 goto retryproc;
1371 }
1372 num_aio_resv_start--;
1373 }
1374 splx(s);
1375done:
1376 return error;
1377}
1378
1379/*
1380 * This routine queues an AIO request, checking for quotas.
1381 */
1382static int
41c20dac 1383aio_aqueue(struct aiocb *job, int type)
984263bc 1384{
7b95be2a 1385 struct proc *p = curproc;
984263bc
MD
1386 struct kaioinfo *ki;
1387
1388 if (p->p_aioinfo == NULL)
1389 aio_init_aioinfo(p);
1390
1391 if (num_queue_count >= max_queue_count)
1392 return EAGAIN;
1393
1394 ki = p->p_aioinfo;
1395 if (ki->kaio_queue_count >= ki->kaio_qallowed_count)
1396 return EAGAIN;
1397
41c20dac 1398 return _aio_aqueue(job, NULL, type);
984263bc
MD
1399}
1400#endif /* VFS_AIO */
1401
1402/*
1403 * Support the aio_return system call, as a side-effect, kernel resources are
1404 * released.
1405 */
1406int
41c20dac 1407aio_return(struct aio_return_args *uap)
984263bc
MD
1408{
1409#ifndef VFS_AIO
1410 return ENOSYS;
1411#else
41c20dac 1412 struct proc *p = curproc;
984263bc
MD
1413 int s;
1414 long jobref;
1415 struct aiocblist *cb, *ncb;
1416 struct aiocb *ujob;
1417 struct kaioinfo *ki;
1418
1419 ki = p->p_aioinfo;
1420 if (ki == NULL)
1421 return EINVAL;
1422
1423 ujob = uap->aiocbp;
1424
1425 jobref = fuword(&ujob->_aiocb_private.kernelinfo);
1426 if (jobref == -1 || jobref == 0)
1427 return EINVAL;
1428
1429 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1430 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo) ==
1431 jobref) {
1432 if (ujob == cb->uuaiocb) {
90b9818c 1433 uap->lmsg.u.ms_result =
984263bc
MD
1434 cb->uaiocb._aiocb_private.status;
1435 } else
90b9818c 1436 uap->lmsg.u.ms_result = EFAULT;
984263bc
MD
1437 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
1438 p->p_stats->p_ru.ru_oublock +=
1439 cb->outputcharge;
1440 cb->outputcharge = 0;
1441 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
1442 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
1443 cb->inputcharge = 0;
1444 }
1445 aio_free_entry(cb);
1446 return 0;
1447 }
1448 }
1449 s = splbio();
1450 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = ncb) {
1451 ncb = TAILQ_NEXT(cb, plist);
1452 if (((intptr_t) cb->uaiocb._aiocb_private.kernelinfo)
1453 == jobref) {
1454 splx(s);
1455 if (ujob == cb->uuaiocb) {
90b9818c 1456 uap->lmsg.u.ms_result =
984263bc
MD
1457 cb->uaiocb._aiocb_private.status;
1458 } else
90b9818c 1459 uap->lmsg.u.ms_result = EFAULT;
984263bc
MD
1460 aio_free_entry(cb);
1461 return 0;
1462 }
1463 }
1464 splx(s);
1465
1466 return (EINVAL);
1467#endif /* VFS_AIO */
1468}
1469
1470/*
1471 * Allow a process to wakeup when any of the I/O requests are completed.
1472 */
1473int
41c20dac 1474aio_suspend(struct aio_suspend_args *uap)
984263bc
MD
1475{
1476#ifndef VFS_AIO
1477 return ENOSYS;
1478#else
41c20dac 1479 struct proc *p = curproc;
984263bc
MD
1480 struct timeval atv;
1481 struct timespec ts;
1482 struct aiocb *const *cbptr, *cbp;
1483 struct kaioinfo *ki;
1484 struct aiocblist *cb;
1485 int i;
1486 int njoblist;
1487 int error, s, timo;
1488 long *ijoblist;
1489 struct aiocb **ujoblist;
1490
1491 if (uap->nent > AIO_LISTIO_MAX)
1492 return EINVAL;
1493
1494 timo = 0;
1495 if (uap->timeout) {
1496 /* Get timespec struct. */
1497 if ((error = copyin(uap->timeout, &ts, sizeof(ts))) != 0)
1498 return error;
1499
1500 if (ts.tv_nsec < 0 || ts.tv_nsec >= 1000000000)
1501 return (EINVAL);
1502
1503 TIMESPEC_TO_TIMEVAL(&atv, &ts);
1504 if (itimerfix(&atv))
1505 return (EINVAL);
1506 timo = tvtohz(&atv);
1507 }
1508
1509 ki = p->p_aioinfo;
1510 if (ki == NULL)
1511 return EAGAIN;
1512
1513 njoblist = 0;
1514 ijoblist = zalloc(aiol_zone);
1515 ujoblist = zalloc(aiol_zone);
1516 cbptr = uap->aiocbp;
1517
1518 for (i = 0; i < uap->nent; i++) {
1519 cbp = (struct aiocb *)(intptr_t)fuword(&cbptr[i]);
1520 if (cbp == 0)
1521 continue;
1522 ujoblist[njoblist] = cbp;
1523 ijoblist[njoblist] = fuword(&cbp->_aiocb_private.kernelinfo);
1524 njoblist++;
1525 }
1526
1527 if (njoblist == 0) {
1528 zfree(aiol_zone, ijoblist);
1529 zfree(aiol_zone, ujoblist);
1530 return 0;
1531 }
1532
1533 error = 0;
1534 for (;;) {
1535 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1536 for (i = 0; i < njoblist; i++) {
1537 if (((intptr_t)
1538 cb->uaiocb._aiocb_private.kernelinfo) ==
1539 ijoblist[i]) {
1540 if (ujoblist[i] != cb->uuaiocb)
1541 error = EINVAL;
1542 zfree(aiol_zone, ijoblist);
1543 zfree(aiol_zone, ujoblist);
1544 return error;
1545 }
1546 }
1547 }
1548
1549 s = splbio();
1550 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb =
1551 TAILQ_NEXT(cb, plist)) {
1552 for (i = 0; i < njoblist; i++) {
1553 if (((intptr_t)
1554 cb->uaiocb._aiocb_private.kernelinfo) ==
1555 ijoblist[i]) {
1556 splx(s);
1557 if (ujoblist[i] != cb->uuaiocb)
1558 error = EINVAL;
1559 zfree(aiol_zone, ijoblist);
1560 zfree(aiol_zone, ujoblist);
1561 return error;
1562 }
1563 }
1564 }
1565
1566 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1567 error = tsleep(p, PCATCH, "aiospn", timo);
984263bc
MD
1568 splx(s);
1569
1570 if (error == ERESTART || error == EINTR) {
1571 zfree(aiol_zone, ijoblist);
1572 zfree(aiol_zone, ujoblist);
1573 return EINTR;
1574 } else if (error == EWOULDBLOCK) {
1575 zfree(aiol_zone, ijoblist);
1576 zfree(aiol_zone, ujoblist);
1577 return EAGAIN;
1578 }
1579 }
1580
1581/* NOTREACHED */
1582 return EINVAL;
1583#endif /* VFS_AIO */
1584}
1585
1586/*
1587 * aio_cancel cancels any non-physio aio operations not currently in
1588 * progress.
1589 */
1590int
41c20dac 1591aio_cancel(struct aio_cancel_args *uap)
984263bc
MD
1592{
1593#ifndef VFS_AIO
1594 return ENOSYS;
1595#else
41c20dac 1596 struct proc *p = curproc;
984263bc
MD
1597 struct kaioinfo *ki;
1598 struct aiocblist *cbe, *cbn;
1599 struct file *fp;
1600 struct filedesc *fdp;
1601 struct socket *so;
1602 struct proc *po;
1603 int s,error;
1604 int cancelled=0;
1605 int notcancelled=0;
1606 struct vnode *vp;
1607
1608 fdp = p->p_fd;
1609 if ((u_int)uap->fd >= fdp->fd_nfiles ||
1610 (fp = fdp->fd_ofiles[uap->fd]) == NULL)
1611 return (EBADF);
1612
1613 if (fp->f_type == DTYPE_VNODE) {
1614 vp = (struct vnode *)fp->f_data;
1615
1616 if (vn_isdisk(vp,&error)) {
90b9818c 1617 uap->lmsg.u.ms_result = AIO_NOTCANCELED;
984263bc
MD
1618 return 0;
1619 }
1620 } else if (fp->f_type == DTYPE_SOCKET) {
1621 so = (struct socket *)fp->f_data;
1622
1623 s = splnet();
1624
1625 for (cbe = TAILQ_FIRST(&so->so_aiojobq); cbe; cbe = cbn) {
1626 cbn = TAILQ_NEXT(cbe, list);
1627 if ((uap->aiocbp == NULL) ||
1628 (uap->aiocbp == cbe->uuaiocb) ) {
1629 po = cbe->userproc;
1630 ki = po->p_aioinfo;
1631 TAILQ_REMOVE(&so->so_aiojobq, cbe, list);
1632 TAILQ_REMOVE(&ki->kaio_sockqueue, cbe, plist);
1633 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe, plist);
1634 if (ki->kaio_flags & KAIO_WAKEUP) {
1635 wakeup(po);
1636 }
1637 cbe->jobstate = JOBST_JOBFINISHED;
1638 cbe->uaiocb._aiocb_private.status=-1;
1639 cbe->uaiocb._aiocb_private.error=ECANCELED;
1640 cancelled++;
1641/* XXX cancelled, knote? */
1642 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1643 SIGEV_SIGNAL)
1644 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1645 if (uap->aiocbp)
1646 break;
1647 }
1648 }
1649 splx(s);
1650
1651 if ((cancelled) && (uap->aiocbp)) {
90b9818c 1652 uap->lmsg.u.ms_result = AIO_CANCELED;
984263bc
MD
1653 return 0;
1654 }
1655 }
1656 ki=p->p_aioinfo;
1657 if (ki == NULL)
1658 goto done;
1659 s = splnet();
1660
1661 for (cbe = TAILQ_FIRST(&ki->kaio_jobqueue); cbe; cbe = cbn) {
1662 cbn = TAILQ_NEXT(cbe, plist);
1663
1664 if ((uap->fd == cbe->uaiocb.aio_fildes) &&
1665 ((uap->aiocbp == NULL ) ||
1666 (uap->aiocbp == cbe->uuaiocb))) {
1667
1668 if (cbe->jobstate == JOBST_JOBQGLOBAL) {
1669 TAILQ_REMOVE(&aio_jobs, cbe, list);
1670 TAILQ_REMOVE(&ki->kaio_jobqueue, cbe, plist);
1671 TAILQ_INSERT_TAIL(&ki->kaio_jobdone, cbe,
1672 plist);
1673 cancelled++;
1674 ki->kaio_queue_finished_count++;
1675 cbe->jobstate = JOBST_JOBFINISHED;
1676 cbe->uaiocb._aiocb_private.status = -1;
1677 cbe->uaiocb._aiocb_private.error = ECANCELED;
1678/* XXX cancelled, knote? */
1679 if (cbe->uaiocb.aio_sigevent.sigev_notify ==
1680 SIGEV_SIGNAL)
1681 psignal(cbe->userproc, cbe->uaiocb.aio_sigevent.sigev_signo);
1682 } else {
1683 notcancelled++;
1684 }
1685 }
1686 }
1687 splx(s);
1688done:
1689 if (notcancelled) {
90b9818c 1690 uap->lmsg.u.ms_result = AIO_NOTCANCELED;
984263bc
MD
1691 return 0;
1692 }
1693 if (cancelled) {
90b9818c 1694 uap->lmsg.u.ms_result = AIO_CANCELED;
984263bc
MD
1695 return 0;
1696 }
90b9818c 1697 uap->lmsg.u.ms_result = AIO_ALLDONE;
984263bc
MD
1698
1699 return 0;
1700#endif /* VFS_AIO */
1701}
1702
1703/*
1704 * aio_error is implemented in the kernel level for compatibility purposes only.
1705 * For a user mode async implementation, it would be best to do it in a userland
1706 * subroutine.
1707 */
1708int
41c20dac 1709aio_error(struct aio_error_args *uap)
984263bc
MD
1710{
1711#ifndef VFS_AIO
1712 return ENOSYS;
1713#else
41c20dac 1714 struct proc *p = curproc;
984263bc
MD
1715 int s;
1716 struct aiocblist *cb;
1717 struct kaioinfo *ki;
1718 long jobref;
1719
1720 ki = p->p_aioinfo;
1721 if (ki == NULL)
1722 return EINVAL;
1723
1724 jobref = fuword(&uap->aiocbp->_aiocb_private.kernelinfo);
1725 if ((jobref == -1) || (jobref == 0))
1726 return EINVAL;
1727
1728 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1729 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1730 jobref) {
90b9818c 1731 uap->lmsg.u.ms_result = cb->uaiocb._aiocb_private.error;
984263bc
MD
1732 return 0;
1733 }
1734 }
1735
1736 s = splnet();
1737
1738 for (cb = TAILQ_FIRST(&ki->kaio_jobqueue); cb; cb = TAILQ_NEXT(cb,
1739 plist)) {
1740 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1741 jobref) {
90b9818c 1742 uap->lmsg.u.ms_result = EINPROGRESS;
984263bc
MD
1743 splx(s);
1744 return 0;
1745 }
1746 }
1747
1748 for (cb = TAILQ_FIRST(&ki->kaio_sockqueue); cb; cb = TAILQ_NEXT(cb,
1749 plist)) {
1750 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1751 jobref) {
90b9818c 1752 uap->lmsg.u.ms_result = EINPROGRESS;
984263bc
MD
1753 splx(s);
1754 return 0;
1755 }
1756 }
1757 splx(s);
1758
1759 s = splbio();
1760 for (cb = TAILQ_FIRST(&ki->kaio_bufdone); cb; cb = TAILQ_NEXT(cb,
1761 plist)) {
1762 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1763 jobref) {
90b9818c 1764 uap->lmsg.u.ms_result = cb->uaiocb._aiocb_private.error;
984263bc
MD
1765 splx(s);
1766 return 0;
1767 }
1768 }
1769
1770 for (cb = TAILQ_FIRST(&ki->kaio_bufqueue); cb; cb = TAILQ_NEXT(cb,
1771 plist)) {
1772 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo) ==
1773 jobref) {
90b9818c 1774 uap->lmsg.u.ms_result = EINPROGRESS;
984263bc
MD
1775 splx(s);
1776 return 0;
1777 }
1778 }
1779 splx(s);
1780
1781#if (0)
1782 /*
1783 * Hack for lio.
1784 */
1785 status = fuword(&uap->aiocbp->_aiocb_private.status);
1786 if (status == -1)
1787 return fuword(&uap->aiocbp->_aiocb_private.error);
1788#endif
1789 return EINVAL;
1790#endif /* VFS_AIO */
1791}
1792
1793/* syscall - asynchronous read from a file (REALTIME) */
1794int
41c20dac 1795aio_read(struct aio_read_args *uap)
984263bc
MD
1796{
1797#ifndef VFS_AIO
1798 return ENOSYS;
1799#else
41c20dac 1800 return aio_aqueue(uap->aiocbp, LIO_READ);
984263bc
MD
1801#endif /* VFS_AIO */
1802}
1803
1804/* syscall - asynchronous write to a file (REALTIME) */
1805int
41c20dac 1806aio_write(struct aio_write_args *uap)
984263bc
MD
1807{
1808#ifndef VFS_AIO
1809 return ENOSYS;
1810#else
41c20dac 1811 return aio_aqueue(uap->aiocbp, LIO_WRITE);
984263bc
MD
1812#endif /* VFS_AIO */
1813}
1814
1815/* syscall - XXX undocumented */
1816int
41c20dac 1817lio_listio(struct lio_listio_args *uap)
984263bc
MD
1818{
1819#ifndef VFS_AIO
1820 return ENOSYS;
1821#else
41c20dac 1822 struct proc *p = curproc;
984263bc
MD
1823 int nent, nentqueued;
1824 struct aiocb *iocb, * const *cbptr;
1825 struct aiocblist *cb;
1826 struct kaioinfo *ki;
1827 struct aio_liojob *lj;
1828 int error, runningcode;
1829 int nerror;
1830 int i;
1831 int s;
1832
1833 if ((uap->mode != LIO_NOWAIT) && (uap->mode != LIO_WAIT))
1834 return EINVAL;
1835
1836 nent = uap->nent;
1837 if (nent > AIO_LISTIO_MAX)
1838 return EINVAL;
1839
1840 if (p->p_aioinfo == NULL)
1841 aio_init_aioinfo(p);
1842
1843 if ((nent + num_queue_count) > max_queue_count)
1844 return EAGAIN;
1845
1846 ki = p->p_aioinfo;
1847 if ((nent + ki->kaio_queue_count) > ki->kaio_qallowed_count)
1848 return EAGAIN;
1849
1850 lj = zalloc(aiolio_zone);
1851 if (!lj)
1852 return EAGAIN;
1853
1854 lj->lioj_flags = 0;
1855 lj->lioj_buffer_count = 0;
1856 lj->lioj_buffer_finished_count = 0;
1857 lj->lioj_queue_count = 0;
1858 lj->lioj_queue_finished_count = 0;
1859 lj->lioj_ki = ki;
1860
1861 /*
1862 * Setup signal.
1863 */
1864 if (uap->sig && (uap->mode == LIO_NOWAIT)) {
1865 error = copyin(uap->sig, &lj->lioj_signal,
1866 sizeof(lj->lioj_signal));
1867 if (error) {
1868 zfree(aiolio_zone, lj);
1869 return error;
1870 }
1871 if (!_SIG_VALID(lj->lioj_signal.sigev_signo)) {
1872 zfree(aiolio_zone, lj);
1873 return EINVAL;
1874 }
1875 lj->lioj_flags |= LIOJ_SIGNAL;
1876 lj->lioj_flags &= ~LIOJ_SIGNAL_POSTED;
1877 } else
1878 lj->lioj_flags &= ~LIOJ_SIGNAL;
1879
1880 TAILQ_INSERT_TAIL(&ki->kaio_liojoblist, lj, lioj_list);
1881 /*
1882 * Get pointers to the list of I/O requests.
1883 */
1884 nerror = 0;
1885 nentqueued = 0;
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)) {
41c20dac 1890 error = _aio_aqueue(iocb, lj, 0);
984263bc
MD
1891 if (error == 0)
1892 nentqueued++;
1893 else
1894 nerror++;
1895 }
1896 }
1897
1898 /*
1899 * If we haven't queued any, then just return error.
1900 */
1901 if (nentqueued == 0)
1902 return 0;
1903
1904 /*
1905 * Calculate the appropriate error return.
1906 */
1907 runningcode = 0;
1908 if (nerror)
1909 runningcode = EIO;
1910
1911 if (uap->mode == LIO_WAIT) {
1912 int command, found, jobref;
1913
1914 for (;;) {
1915 found = 0;
1916 for (i = 0; i < uap->nent; i++) {
1917 /*
1918 * Fetch address of the control buf pointer in
1919 * user space.
1920 */
1921 iocb = (struct aiocb *)
1922 (intptr_t)fuword(&cbptr[i]);
1923 if (((intptr_t)iocb == -1) || ((intptr_t)iocb
1924 == 0))
1925 continue;
1926
1927 /*
1928 * Fetch the associated command from user space.
1929 */
1930 command = fuword(&iocb->aio_lio_opcode);
1931 if (command == LIO_NOP) {
1932 found++;
1933 continue;
1934 }
1935
1936 jobref = fuword(&iocb->_aiocb_private.kernelinfo);
1937
1938 TAILQ_FOREACH(cb, &ki->kaio_jobdone, plist) {
1939 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1940 == jobref) {
1941 if (cb->uaiocb.aio_lio_opcode
1942 == LIO_WRITE) {
1943 p->p_stats->p_ru.ru_oublock
1944 +=
1945 cb->outputcharge;
1946 cb->outputcharge = 0;
1947 } else if (cb->uaiocb.aio_lio_opcode
1948 == LIO_READ) {
1949 p->p_stats->p_ru.ru_inblock
1950 += cb->inputcharge;
1951 cb->inputcharge = 0;
1952 }
1953 found++;
1954 break;
1955 }
1956 }
1957
1958 s = splbio();
1959 TAILQ_FOREACH(cb, &ki->kaio_bufdone, plist) {
1960 if (((intptr_t)cb->uaiocb._aiocb_private.kernelinfo)
1961 == jobref) {
1962 found++;
1963 break;
1964 }
1965 }
1966 splx(s);
1967 }
1968
1969 /*
1970 * If all I/Os have been disposed of, then we can
1971 * return.
1972 */
1973 if (found == nentqueued)
1974 return runningcode;
1975
1976 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 1977 error = tsleep(p, PCATCH, "aiospn", 0);
984263bc
MD
1978
1979 if (error == EINTR)
1980 return EINTR;
1981 else if (error == EWOULDBLOCK)
1982 return EAGAIN;
1983 }
1984 }
1985
1986 return runningcode;
1987#endif /* VFS_AIO */
1988}
1989
1990#ifdef VFS_AIO
1991/*
1992 * This is a weird hack so that we can post a signal. It is safe to do so from
1993 * a timeout routine, but *not* from an interrupt routine.
1994 */
1995static void
1996process_signal(void *aioj)
1997{
1998 struct aiocblist *aiocbe = aioj;
1999 struct aio_liojob *lj = aiocbe->lio;
2000 struct aiocb *cb = &aiocbe->uaiocb;
2001
2002 if ((lj) && (lj->lioj_signal.sigev_notify == SIGEV_SIGNAL) &&
2003 (lj->lioj_queue_count == lj->lioj_queue_finished_count)) {
2004 psignal(lj->lioj_ki->kaio_p, lj->lioj_signal.sigev_signo);
2005 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2006 }
2007
2008 if (cb->aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2009 psignal(aiocbe->userproc, cb->aio_sigevent.sigev_signo);
2010}
2011
2012/*
2013 * Interrupt handler for physio, performs the necessary process wakeups, and
2014 * signals.
2015 */
2016static void
2017aio_physwakeup(struct buf *bp)
2018{
2019 struct aiocblist *aiocbe;
2020 struct proc *p;
2021 struct kaioinfo *ki;
2022 struct aio_liojob *lj;
2023
2024 wakeup(bp);
2025
2026 aiocbe = (struct aiocblist *)bp->b_spc;
2027 if (aiocbe) {
2028 p = bp->b_caller1;
2029
2030 aiocbe->jobstate = JOBST_JOBBFINISHED;
2031 aiocbe->uaiocb._aiocb_private.status -= bp->b_resid;
2032 aiocbe->uaiocb._aiocb_private.error = 0;
2033 aiocbe->jobflags |= AIOCBLIST_DONE;
2034
2035 if (bp->b_flags & B_ERROR)
2036 aiocbe->uaiocb._aiocb_private.error = bp->b_error;
2037
2038 lj = aiocbe->lio;
2039 if (lj) {
2040 lj->lioj_buffer_finished_count++;
2041
2042 /*
2043 * wakeup/signal if all of the interrupt jobs are done.
2044 */
2045 if (lj->lioj_buffer_finished_count ==
2046 lj->lioj_buffer_count) {
2047 /*
2048 * Post a signal if it is called for.
2049 */
2050 if ((lj->lioj_flags &
2051 (LIOJ_SIGNAL|LIOJ_SIGNAL_POSTED)) ==
2052 LIOJ_SIGNAL) {
2053 lj->lioj_flags |= LIOJ_SIGNAL_POSTED;
2054 aiocbe->timeouthandle =
2055 timeout(process_signal,
2056 aiocbe, 0);
2057 }
2058 }
2059 }
2060
2061 ki = p->p_aioinfo;
2062 if (ki) {
2063 ki->kaio_buffer_finished_count++;
2064 TAILQ_REMOVE(&aio_bufjobs, aiocbe, list);
2065 TAILQ_REMOVE(&ki->kaio_bufqueue, aiocbe, plist);
2066 TAILQ_INSERT_TAIL(&ki->kaio_bufdone, aiocbe, plist);
2067
2068 KNOTE(&aiocbe->klist, 0);
2069 /* Do the wakeup. */
2070 if (ki->kaio_flags & (KAIO_RUNDOWN|KAIO_WAKEUP)) {
2071 ki->kaio_flags &= ~KAIO_WAKEUP;
2072 wakeup(p);
2073 }
2074 }
2075
2076 if (aiocbe->uaiocb.aio_sigevent.sigev_notify == SIGEV_SIGNAL)
2077 aiocbe->timeouthandle =
2078 timeout(process_signal, aiocbe, 0);
2079 }
2080}
2081#endif /* VFS_AIO */
2082
2083/* syscall - wait for the next completion of an aio request */
2084int
41c20dac 2085aio_waitcomplete(struct aio_waitcomplete_args *uap)
984263bc
MD
2086{
2087#ifndef VFS_AIO
2088 return ENOSYS;
2089#else
41c20dac 2090 struct proc *p = curproc;
984263bc
MD
2091 struct timeval atv;
2092 struct timespec ts;
2093 struct kaioinfo *ki;
2094 struct aiocblist *cb = NULL;
2095 int error, s, timo;
2096
2097 suword(uap->aiocbp, (int)NULL);
2098
2099 timo = 0;
2100 if (uap->timeout) {
2101 /* Get timespec struct. */
2102 error = copyin(uap->timeout, &ts, sizeof(ts));
2103 if (error)
2104 return error;
2105
2106 if ((ts.tv_nsec < 0) || (ts.tv_nsec >= 1000000000))
2107 return (EINVAL);
2108
2109 TIMESPEC_TO_TIMEVAL(&atv, &ts);
2110 if (itimerfix(&atv))
2111 return (EINVAL);
2112 timo = tvtohz(&atv);
2113 }
2114
2115 ki = p->p_aioinfo;
2116 if (ki == NULL)
2117 return EAGAIN;
2118
2119 for (;;) {
2120 if ((cb = TAILQ_FIRST(&ki->kaio_jobdone)) != 0) {
2121 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
90b9818c 2122 uap->lmsg.u.ms_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2123 if (cb->uaiocb.aio_lio_opcode == LIO_WRITE) {
2124 p->p_stats->p_ru.ru_oublock +=
2125 cb->outputcharge;
2126 cb->outputcharge = 0;
2127 } else if (cb->uaiocb.aio_lio_opcode == LIO_READ) {
2128 p->p_stats->p_ru.ru_inblock += cb->inputcharge;
2129 cb->inputcharge = 0;
2130 }
2131 aio_free_entry(cb);
2132 return cb->uaiocb._aiocb_private.error;
2133 }
2134
2135 s = splbio();
2136 if ((cb = TAILQ_FIRST(&ki->kaio_bufdone)) != 0 ) {
2137 splx(s);
2138 suword(uap->aiocbp, (uintptr_t)cb->uuaiocb);
90b9818c 2139 uap->lmsg.u.ms_result = cb->uaiocb._aiocb_private.status;
984263bc
MD
2140 aio_free_entry(cb);
2141 return cb->uaiocb._aiocb_private.error;
2142 }
2143
2144 ki->kaio_flags |= KAIO_WAKEUP;
377d4740 2145 error = tsleep(p, PCATCH, "aiowc", timo);
984263bc
MD
2146 splx(s);
2147
2148 if (error == ERESTART)
2149 return EINTR;
2150 else if (error < 0)
2151 return error;
2152 else if (error == EINTR)
2153 return EINTR;
2154 else if (error == EWOULDBLOCK)
2155 return EAGAIN;
2156 }
2157#endif /* VFS_AIO */
2158}
2159
2160#ifndef VFS_AIO
2161static int
2162filt_aioattach(struct knote *kn)
2163{
2164
2165 return (ENXIO);
2166}
2167
2168struct filterops aio_filtops =
2169 { 0, filt_aioattach, NULL, NULL };
2170
2171#else
2172/* kqueue attach function */
2173static int
2174filt_aioattach(struct knote *kn)
2175{
2176 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2177
2178 /*
2179 * The aiocbe pointer must be validated before using it, so
2180 * registration is restricted to the kernel; the user cannot
2181 * set EV_FLAG1.
2182 */
2183 if ((kn->kn_flags & EV_FLAG1) == 0)
2184 return (EPERM);
2185 kn->kn_flags &= ~EV_FLAG1;
2186
2187 SLIST_INSERT_HEAD(&aiocbe->klist, kn, kn_selnext);
2188
2189 return (0);
2190}
2191
2192/* kqueue detach function */
2193static void
2194filt_aiodetach(struct knote *kn)
2195{
2196 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2197
2198 SLIST_REMOVE(&aiocbe->klist, kn, knote, kn_selnext);
2199}
2200
2201/* kqueue filter function */
2202/*ARGSUSED*/
2203static int
2204filt_aio(struct knote *kn, long hint)
2205{
2206 struct aiocblist *aiocbe = (struct aiocblist *)kn->kn_sdata;
2207
2208 kn->kn_data = aiocbe->uaiocb._aiocb_private.error;
2209 if (aiocbe->jobstate != JOBST_JOBFINISHED &&
2210 aiocbe->jobstate != JOBST_JOBBFINISHED)
2211 return (0);
2212 kn->kn_flags |= EV_EOF;
2213 return (1);
2214}
2215
2216struct filterops aio_filtops =
2217 { 0, filt_aioattach, filt_aiodetach, filt_aio };
2218#endif /* VFS_AIO */